AAyano/uf850-vr-teleop-place-corn-in-bowl-raw

UF850 Quest VR Place Corn in Bowl — Raw Data

Raw, untrimmed UF850 teleoperation trajectories collected with a Quest VR controller and two Intel RealSense cameras. No trajectory trimming, resampling, or action normalization has been applied. Camera frames were captured as YUYV and stored as JPEG at collection time.

Task and prompt

Every episode uses the same English task instruction:

Place the corn in the orange bowl.

The canonical prompt is stored in episode_meta.json as task. It is intentionally not duplicated in every step-level data.json.

Task layout

In this external-camera view, the orange bowl is near the center of the workspace and the corn is to its right. The UF850 base is behind the objects.

Dataset summary

Item	Value
Episodes	51 (episode_000–episode_050)
Transition steps	13,235
JPEG images	26,470 (13,235 per camera)
Nominal transition rate	10 Hz
Nominal transition time	22.06 minutes
Image size	1920×1080

Cameras

File	Role	RealSense serial	Capture mode
cam_0.jpg	wrist	845112070404	1920×1080 YUYV @ 30 FPS
cam_1.jpg	external	215222078407	1920×1080 YUYV @ 30 FPS

File layout

To keep the Hub repository reliable and easy to download, each episode is distributed as one uncompressed tar archive. The archive is only a container: the JPEG and JSON file bytes inside are unchanged from the validated raw capture.

README.md
layout_external.jpg
SHA256SUMS
episodes/
  episode_000.tar
  episode_001.tar
  ...

Extract an episode with, for example:

tar -xf episodes/episode_000.tar

Each archive expands to the original capture layout:

episode_NNN/
  episode_meta.json
  step_00000/
    cam_0.jpg
    cam_1.jpg
    data.json
  step_00001/
    ...

Each step contains the synchronized observation and camera pair at time t, plus the raw robot-state difference to the following sample:

action[t] = state[t + 1] - state[t]

Actions are not divided by dt.

data.json

• observations.ee_pos: [x, y, z, roll, pitch, yaw]
• observations.joint_pos: six joint angles
• observations.gripper_pos: raw xArm gripper position
• action.delta_ee_pos: next-minus-current TCP delta
• action.delta_joint_pos: next-minus-current joint delta
• action.delta_gripper: next-minus-current gripper delta
• meta: schema version, step number, sampling time, real dt, camera timestamps/frame numbers/skew, frame age, state-read duration, and wrapped Euler-angle delta

Units are millimetres for TCP XYZ, degrees for TCP Euler angles and joint angles, and the raw xArm scale (nominally 0–850) for the gripper. Delta fields use the corresponding source units.

episode_meta.json

Episode metadata records the task prompt, start/end times, termination reason, camera identities, collection configuration snapshot, step count, and error statistics.

Validation and raw-data notes

• All 51 episodes were checked for contiguous numbering, required files, full 1920×1080 JPEG decoding, finite JSON values, next-current action alignment, camera synchronization, frame ordering, and episode metadata consistency.
• episode_010 was missing episode-level metadata. It was reconstructed from the preserved step timestamps, finalized-file time, shared task, and the collection configuration used by the surrounding episodes.
• A hidden temporary step in episode_010 contained only an incomplete camera write. It was never a committed sample and was removed during cleanup; all published step directories are complete.
• Raw Euler deltas retain the legacy direct-subtraction behavior, including wrap-boundary values near ±360°. A wrapped rotation delta is also available in each step's metadata.