real_workbench

The four tasks of the real-robot Workbench Manipulation cell in one LeRobot v3.0 dataset, pre-processed for π0.5 co-training: 320 episodes (4 x 80), 133,910 frames at 10 fps, single-arm Franka Research 3, two camera views stored at 224x126, actions in the delta_eef space.

task_index	task	episodes	frames	source release
0	Object Identification	80	36,239	Myungkyu/object_retrieval
1	Layout Reconstruction	80	37,559	Myungkyu/layout_reconstruction
2	Movement Reversal	80	31,543	Myungkyu/movement_reversal
3	Object Classification	80	28,569	Myungkyu/object_classification

The labelled companion Myungkyu/real_workbench-preset-gemini holds the same frames with a dense per-frame subtask label instead of the task instruction.

Pre-processing

Both differences from the per-task releases exist to make the stored frame exactly what the policy sees:

• Exterior view cropped. The bottom-right 80 % x 80 % of the 1280x720 capture is kept: the left fifth of the width and the top fifth of the height are dropped. Those regions hold the cell wall, the upper links of the arm and a neighbouring workstation; the workbench itself (far board, center tray, drawer compartments, right container) stays inside the crop, and the objects gain 25 % in linear size. The wrist view is never cropped.
• 224x126 frames. π0.5 resizes any input with resize_with_pad to 224x224 without distortion, which for a 16:9 frame means 224x126 content plus 49-px black bars. Storing 224x126 hands the model its own input resolution, so no rescaling happens at training time. The crop is exactly 16:9, so cropped and uncropped views land on the same size.

Frames are otherwise the source episodes unchanged: same capture, same 10 fps grid, one frame per control step, no temporal resampling. The scaling filter is swscale bilinear in both views.

Layout and features

data/chunk-000/file-*.parquet        packed episode frames (v3.0)
videos/observation.image.exterior_camera/chunk-000/file-*.mp4   exterior, cropped, 224x126
videos/observation.image.wrist_camera/chunk-000/file-*.mp4      wrist, 224x126
meta/{info,stats,tasks,episodes}     v3.0 metadata; stats carries q01..q99 for state and action

Per-frame features: observation.state (8: 7 joint angles + gripper), action (7: the DROID cartesian-velocity command + gripper, the delta_eef convention of the source releases), subtask (here the task instruction; the labelled companion puts a per-frame subtask here), timestamp, frame_index, episode_index, index, task_index, and the two views.

Task instructions

• Object Identification — Collect every object that matches those shown inside the reference tray, place them into the collection box, and close the lid.
• Layout Reconstruction — Memorize the object layout on the workspace, transfer the objects onto the layout board, and then reconstruct the original layout.
• Movement Reversal — Observe the order in which the demonstrator moves the objects, then reverse the sequence to restore them to their original positions.
• Object Classification — Place each object from the tray into the drawer, board, or container that contains its matching object.

Source

DROID-style teleoperation captures (Meta Quest controller, 10 Hz control loop, two ZED cameras), Apache-2.0. This dataset re-encodes and merges the four public per-task releases listed above; state and action are unchanged.