dexhand-midtrain

Simulated mid-training data for a dual Marvin M6-S (7-DoF) + Wuji 20-DoF hand robot (54-D state / action), generated in RoboDojo (Isaac Sim) with the robot model v2 (2026-10-22: right hand on the mirror mount, right-arm home = joint mirror of the left-arm home, so sim right-arm joints map to the real right arm by the sign vector [-1, 1, -1, 1, -1, 1, -1]). Head camera: RealSense D435 model (640x480), wrist cameras D405. Each batch is split into LeRobot shards (lerobot/<batch>_<tag>_p0_sNNN/), cut into atomic sub-task segments with language instructions; every shard was verified against the generator's HDF5 (state / action / task / index: 0 mismatches, image PSNR >= 33 dB) before upload. Every episode's meta carries robot: ... model v2.

batch	task	episodes (success / attempts)	shards	notes
cover_m2	cover_blocks	100 / 114	5 x 140 segments (7 per episode)	appearance randomisation, arm noise, 15 % injected faults
align_v2	align_blocks	103 / 249	5 x 40	both arms push the ruler
bottles_v2	put_bottles_into_dustbin	101 / 124	5 x 100	4 bottles on the left half, left arm
blocks_v2	stack_blocks	101 / 389	5 x 60	3 / 3.5 / 4 cm cubes, thumb-index pinch, either half
bowls_v2	stack_bowls	106 / 190	5 x 60	11 cm bowls, three on one half, one arm

Generator and tooling: github (private) dexhand, commit history in each shard's meta. More tasks follow as their batches are verified.