--- license: apache-2.0 task_categories: - robotics tags: - LeRobot - robotics - franka - isaac-sim - cube-stacking - motion-planner - gepa - simulation - not-teleop - reset-left configs: - config_name: default data_files: data/*/*.parquet --- This dataset was created using [LeRobot](https://github.com/huggingface/lerobot). ## Dataset Description > ## ⚠️ Read before mixing this with other DecentVLA data > > **This is a `Franka Panda` in simulation, driven by a motion planner.** > Every other dataset in this organisation is **SO-101 hardware, teleoperated by a > human**. Two things therefore differ, and neither is cosmetic: > > - **The embodiment.** A 7-DoF Franka arm, not a 5-DoF SO-101. Different > kinematics, different joint count, different workspace. > - **The action space.** `action` here is 8-D — 7 arm joints plus a gripper — and > its values are normalised joint targets emitted by a planner. It is not the > same quantity as an SO-101 teleop action and the two are not interchangeable. > > Concatenating these with the SO-101 sets as if they shared an embodiment will > train on incompatible action vectors. That is worth checking deliberately rather > than discovering in a loss curve. **Red cube on Blue cube**, Franka Panda in Isaac Sim, reset distribution **left**. > `Put the red cube on top of the blue cube.` ### How the actions were produced **These are not teleoperated demonstrations.** Every action is the output of a **motion planner discovered by GEPA**, an LLM-driven program search, and then frozen: the planner named `cubestack_fast` was promoted at **1.026 train / 0.937 held-out** and was not tuned, retrained or edited to produce this dataset. It is the same planner across all twelve datasets in this collection, which is the point — the datasets differ in the scene, never in the policy that acted in it. ### The reset distribution The two cubes spawn in the **+y half** of the region each one's ROLE was searched over, so the whole episode is displaced toward **the arm's left** relative to its twin. The halves are subsets of the region the planner was evolved on, which is what lets it be recorded rather than re-tuned. | role | colour | spawn box, metres (x, y, z) | |---|---|---| | moved | red | [0.42, -0.11, 0.0235] to [0.58, -0.06, 0.0235] | | base | blue | [0.42, 0.11, 0.0235] to [0.58, 0.16, 0.0235] | `observation.environment_state` carries both cubes as `(xyz + quaternion)` **in the robot base frame**, ordered `['moved', 'base']`. Note that the moved cube's y converges toward the base cube's over an episode — it is *carried onto it*. The reset distribution is the first frame of each episode; that is what distinguishes this dataset from its twin. ### Contents | | | |---|---| | episodes | **100** | | frames | 31,450 | | fps | 15 | | success rate | **1.000** (100 of 100 attempts) | | cameras | `observation.images.overhead`, `observation.images.wrist`, 480x640 | | state / action | 8-D — 7 arm joints + gripper | | scene | `cubestack_red_on_blue_yplus` (`71039905dba43067`) | **Successful episodes only.** Failed attempts are not included and are not padding: the count above is the number of episodes in which the cube was placed and stayed placed after the arm returned home. - **Homepage:** [More Information Needed] - **Paper:** [More Information Needed] - **License:** apache-2.0 ## Dataset Structure [meta/info.json](meta/info.json): ```json { "codebase_version": "v3.0", "fps": 15, "features": { "observation.images.overhead": { "dtype": "video", "shape": [ 480, 640, 3 ], "names": [ "height", "width", "channels" ], "info": { "video.height": 480, "video.width": 640, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.fps": 15, "video.channels": 3, "has_audio": false, "video.g": 2, "video.crf": 30, "video.preset": 12, "video.fast_decode": 0, "video.video_backend": "pyav", "video.extra_options": {}, "is_depth_map": false } }, "observation.images.wrist": { "dtype": "video", "shape": [ 480, 640, 3 ], "names": [ "height", "width", "channels" ], "info": { "video.height": 480, "video.width": 640, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.fps": 15, "video.channels": 3, "has_audio": false, "video.g": 2, "video.crf": 30, "video.preset": 12, "video.fast_decode": 0, "video.video_backend": "pyav", "video.extra_options": {}, "is_depth_map": false } }, "observation.state": { "dtype": "float32", "shape": [ 8 ], "names": [ "panda_joint1", "panda_joint2", "panda_joint3", "panda_joint4", "panda_joint5", "panda_joint6", "panda_joint7", "gripper" ] }, "observation.environment_state": { "dtype": "float32", "shape": [ 14 ], "names": [ "moved.x", "moved.y", "moved.z", "moved.qw", "moved.qx", "moved.qy", "moved.qz", "base.x", "base.y", "base.z", "base.qw", "base.qx", "base.qy", "base.qz" ] }, "action": { "dtype": "float32", "shape": [ 8 ], "names": [ "panda_joint1", "panda_joint2", "panda_joint3", "panda_joint4", "panda_joint5", "panda_joint6", "panda_joint7", "gripper" ] }, "timestamp": { "dtype": "float32", "shape": [ 1 ], "names": null }, "frame_index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "episode_index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "task_index": { "dtype": "int64", "shape": [ 1 ], "names": null } }, "total_episodes": 100, "total_frames": 31450, "total_tasks": 1, "chunks_size": 1000, "data_files_size_in_mb": 100, "video_files_size_in_mb": 200, "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet", "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4", "robot_type": "franka", "splits": { "train": "0:100" } } ``` ## Citation **BibTeX:** ```bibtex [More Information Needed] ```