so101_chess_corrections

One thousand expert corrections for an SO-101 arm capturing chess pieces: episodes that begin where a trained policy had just gone wrong, and show the way out. They are the data behind so101_chess_molmoact2_dagger, which took captures at the shipped 30-action horizon from 68 to 78 of 100 on the same positions. Recorded in MuJoCo. The demonstrations they extend are so101_chess.

episodes	1,000, all captures (take the piece on d5 off the board)
frames	89,862 at 10 Hz
cameras	observation.images.top (overhead) and observation.images.wrist, 640 x 480
state and action	six joints, SO-101 degrees; the action is the expert's commanded target
format	LeRobot v3.0

How a correction is recorded

so101_chess_molmoact2 plays a capture, executing all 30 actions of each predicted chunk - three seconds of motion decided before it starts. Every 15 control steps a rule looks at the board. If the attempt is going wrong, the episode is handed to the scripted expert, which finishes the capture from wherever the arm is. Only the expert's part is recorded; the policy's own behaviour is not in the data.

what triggered the hand-over	episodes
the named piece was still not lifted at 40% of the step budget	835
the piece was pushed across its square without being lifted	136
a neighbouring piece was disturbed	29

A piece the policy had already knocked over ends the episode with nothing recorded - the expert has no grasp for a piece on its side. 6,056 policy episodes gave 1,524 hand-overs; these are the 1,000 the expert completed, verified like every demonstration: the piece in the tray, upright, nothing else displaced.

scenes.jsonl has one line per episode, in order: the instruction, the trigger (recovery), and the board shift of that episode.

Using it

from lerobot.datasets.lerobot_dataset import LeRobotDataset
corrections = LeRobotDataset("XvKuoMing/so101_chess_corrections")

The published model counted each correction three times beside 8,083 clean demonstrations (tools/merge_datasets.py --repeat in the project repository): a thousand episodes are a small share of the set, and the weighting is part of the result. 1,982 of those clean demonstrations were extra captures recorded for an earlier experiment; they changed nothing on their own and are not published separately.

Limits

Simulation only. The corrections come from one policy's failures on one rig; a different policy fails elsewhere and wants its own. docs/EXPERIMENTS.md in the project repository, sections 4.9 and 4.10, has the full account.