Simulated SO-101 + TurtleBot 4 handover: all categories combined (2026-10-29, with evening top-up)
One LeRobot v3.0 dataset (tag v3.0) with every validated simulated handover episode recorded so far:
30 fps, robot_type so101_tb4, six AV1 640x360 video streams (camera_left, camera_right, camera_wrist,
camera_d455, camera_front, depth), observation.state (11) and action (8), the same features as the real
teleop dataset MoAIBo/giver_receiver_teleop_merged_1k.
5,790 episodes, 6,079,500 frames (56.3 h). Training set: 5,160 episodes with yaseen_keep_20261029 == True.
Category (dataset_category) |
Episodes | Successful | Notes |
|---|---|---|---|
two_robot_regular |
3,024 | 3,024 | side-by-side, opposite pass, catch-up, same direction; 42 trials per (cell, role) |
two_robot_L |
1,008 | 1,008 | L-handover with random starts |
relay_3_robot |
654 | 654 | 218 kept 3-robot relay runs x 3 robots |
two_robot_rl_demo |
1,104 | 476 | 552 trials (giver + receiver) from the RL residual policy on YCB household objects; all recorded attempts, success flagged |
All episodes use the corrected camera rig: wrist mount A, wrist fisheye, side-camera tilt from real data, D455 depth invalid band. Episodes 0-4031 are byte-identical to the two-robot-only version and 0-5375 to the morning combined build.
Per-episode labels:
meta/episode_categories.parquet(episode_index,dataset_category,source_dataset,success,yaseen_keep_20261029): train onyaseen_keep_20261029 == True(successes only, minus one reviewed trial, episodes 530/531).meta/sim_episodes.parquet: the source sidecar columns (handover_type, cell, role, object, DR settings, ...) plusdataset_category,source_datasetandsource_episode_index.meta/sim_provenance.json,meta/COMBINED_SUMMARY.json.
Earlier versions: git tags two_robot_only_20261026 (4,032 episodes) and combined_5376_20261029 (5,376 episodes).
Not included yet: 23 relay runs recorded before wrist mount A (Oct 21-23); they will be added after their wrist re-render.
- Downloads last month
- 522