real_workbench-taco-gemini

The four tasks of the real-robot Workbench Manipulation cell in one LeRobot v3.0 dataset with reviewed dense high-level labels, 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. Same frames as Myungkyu/real_workbench, with a per-frame subtask label in place of the task instruction.

This is the reviewed companion of real_workbench-preset-gemini. Gemini 3.7 Flash first labelled every episode from the full task context (the subtask catalogue with its DONE WHEN criteria, the sequence rule and the tick rules); the benchmark author then reviewed every tick frame by frame and corrected it. For Movement Reversal and Object Classification the correction was made against the cell's own layout design, which fixes for each episode exactly which subtasks occur and in which order, so those two tasks now carry the designed composition rather than a reading of the video.

task_index	task	episodes	distinct subtask labels
0	Object Identification (instance_retrieval)	80	9
1	Layout Reconstruction (layout_reconstruction)	80	66
2	Movement Reversal (movement_reversal)	80	29
3	Object Classification (object_classification)	80	16

Layout Reconstruction names the destination of each returning object as a cell of a 4 x 4 grid on the center tray, cell (column, row) in the robot's frame — the robot stands to the right of the workbench facing left, so column 1 is its leftmost column and row 1 the row nearest it; cell (1, 1) is the near-left corner and (4, 4) the far-right one. The grid comes from the four corners of the tray measured on the capture-resolution frame; the tray does not move between episodes (corner spread: 90th percentile 1.4-2.1 px over all 80 episodes).

The workbench

Four fixtures, named the same way in every task: the center tray (the square plate in the middle), the left drawer (two compartments, near and far), the right container (the lidded box) and the far board (the stand at the back, left and right half). Four parts: the white holder, the metal ring, the black bracket and the black round base.

Frames and actions

The exterior view is cropped to the bottom-right 80 % x 80 % of the 1280x720 capture (the left fifth of the width and the top fifth of the height are dropped) and stored at 224x126; the wrist view is uncropped at 224x126. π0.5's resize_with_pad turns 224x126 into 224x224 with 49-px bars and no rescaling, so the stored frame is exactly the policy input. action is 7-D delta_eef (6-D cartesian velocity + gripper); observation.state is the 8-D joint vector (7 joints + gripper). There is no keyframe stream: these labels are used without visual memory.

Source

Episodes: Myungkyu/object_retrieval, layout_reconstruction, movement_reversal, object_classification — DROID-style teleoperation captures, Meta Quest controller, cartesian-velocity control at 10 Hz, two ZED cameras, Apache-2.0. This dataset adds the labels and the pre-processing only.