Bigenlight/cube_stack_lerobot_v3

Cube stack — lerobot_v3

Task: stack the red cube on the blue cube

UR7e/GELLO cube-stacking recordings collected on 2026-09-19. Place the red cube on top of the blue cube.

This preview is extracted from take_01_20260919_212809/cam1.mp4; it is not a separate setup photograph or an outcome label.

Contents

60 episodes / 20,177 cam1 master frames at 30 FPS (about 11.2 minutes). Two 1280×720 RGB cameras (cam1 scene, cam2 wrist); no depth. State/action each contain six joints plus gripper. Raw files remain unchanged under raw/<original-take-name>/ in the raw repository.

Collection tags

Exact suffixes are preserved in folder_tag and folder_labels in meta/source_takes.json. recovery_label="recover" selects all recovery variants, consistent with previous lighting releases. Counts: {"untagged": 48, "recover": 9, "recover_many": 3}. Original full take names are retained, including two different takes numbered 01. No take was excluded.

Tags are collection annotations, not verified success/failure labels. No paper train/validation/test partition is assigned; the default LeRobot train range exposes the full corpus.

Conversion

Existing fixed-recorder recipe (convert_takes_to_lerobot.py, LeRobot 0.6.1/v3.0): camera/joint header stamp_s alignment, command/gripper callback t_rel_s, nearest timestamps, cam1 master grid. No servo-lag subtraction or frame dropping. State: ur_q1..6 + grip_pos; action: cmd1..6 + grip_cmd. GELLO leader coordinates are excluded from policy observations.

Companion: Bigenlight/cube_stack_raw.

Independent conversion validation: {'pass': 43, 'fail': 0, 'warn': 0, 'skip': 0, 'total': 43}. This validates conversion fidelity, not robot policy success.