jezv/takeitback-tray

TakeItBack Tray — LeRobot v3.0, v5

1000 successful demonstrations of picking up a cup from the counter and placing it on a tray. This revision replaces the previous 1200-episode dataset on main.

• Environment: MyRoboCasa_TakeItBackTray-v1
• Robot: ds_fetch; control mode: pd_joint_pos
• Collection: 20 Hz; LeRobot dataset: 10 Hz, stride 2
• 1000 episodes, 433108 frames, 3000 RGB videos
• Actions: 13D; observation.state: 12D; global_state: 3D
• Cameras: fetch_hand 128×128, left_base_camera_link and right_base_camera_link 256×256
• Checker: v5; 100 held-out validation seeds, disjoint from collection seeds

Provenance and validation

collection_provenance.json, dataset_manifest.json, and validation_seeds.json describe the collection. Collection retained 1000 trajectories from 1035 attempts: 9 planner failures and 26 safety-checker rejections. The 100 held-out seeds passed planner/checker validation (100/104 candidates).

The action-replay check has not been run on this v5 dataset. The 20/20 result from the superseded 1200-episode revision did not transfer and is not included as validation evidence.

Known limitations

This dataset does not pass the full DATASET_CHECKLIST.md audit. In particular:

• Target visibility is below 95% in 871/1000 episodes (mean 77.15%, minimum 54.8%).
• Gripper actions include intermediate values; they are not strictly binary. Head-pan action has near-zero variance.
• 47.4% of episodes contain backward base motion. Identical adjacent action pairs are 26.86%; longest run is 193 frames at 10 Hz.
• Three episodes contain a roll-joint value outside [-π, π].
• Exact PaliGemma instruction tokenization, LeRobot loader compatibility, video-vs-clean-render error, and full human video review remain unverified.

Use with these limitations in mind; do not treat planner success as evidence that every observation/action-quality criterion passed.