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.

Agreed mid-training task list: cover_blocks, align_blocks, put_bottles_into_dustbin, stack_blocks, stack_bowls, fill_pen_holder, general_pickup (substitute for pour_liquid_into_cup / plug_in_charger, which are not yet solved). pack_v2 and dolls_v2 are extra batches outside that list.

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
pickup_v2	general_pickup	107 / 150	5 x 40 (one pick-hold-place per episode)	household target among the official clutter: cup, yogurt, noodles, glue, toy, sachet; "pick up the "
pack_v2	pack_objects_into_box	102 / 187 (shoe in the box = success)	6 shards (5 x 20 + 2 episodes), 2 segments per episode: "put the shoe in the box" + home	extra data, not in the mid-training task list (produced while the agreed tasks pour / charger were blocked); shoes only (car / toothbrush claws do not lift, hammer pinch fails IK on this hand); shoe on the right half x 0.30-0.45, y <= -0.15
dolls_v2	sort_nesting_dolls_by_size	30 / 220 full episodes (capped at 220 attempts)	11 shards, successful "put the doll in its place" / "bring ... to the middle" / "move ... out of the way" segments, also from failed episodes	extra data, not in the mid-training task list (produced while the agreed tasks pour / charger were blocked); five matryoshka dolls (texture0), per-model 5-finger claws, sorted smallest-left to largest-right
pickup_v3	general_pickup	100 / 127	5 x 40 (one pick-hold-place per episode)	second object set: bottle, figurine, bread, speaker
pens_v2	fill_pen_holder	100 / 211 (two markers standing = success)	11 shards, 388 "put a pen into the pen holder" segments (successful inserts of failed episodes included)	two flat markers per episode, holder spawned on the left (no relay); thumb vs index+middle pinch at the marker's middle, rolled upright, carried above the rim, inserted side by side; no arm noise

Post-training batches (not mid-training data)

Task-specific data for the two post-training tasks, in the same format (sub-task segments). Produced after the mid-training set was frozen; listed here because they live in the same repo (lerobot/kong_v1_*, lerobot/write_v1_*).

batch	task	episodes (success / attempts)	shards	notes
kong_v1	make_kong (simplified)	in production (target 200)	rolling, 20 attempts per shard, "declare a kong: lay down the matching tiles" segments (one per tile; successful tiles of failed episodes included)	official RoboDojo make_kong eval layouts + a cue tile lying in front of the row in place of the opponent's discard; the arm on the tile's side pushes the three matching tiles over from the front with a straight index finger (face up); the other tiles must stay standing. The official second half (standing the discard at the end of the row, needs a hand-over) is not done. Arm noise 0.02 rad on 30 % of episodes
write_v1	write_on_paper (handwriting, as in arXiv 2609.19666 "Write on Paper": the word WUJI)	in production (target 200)	rolling, 20 attempts per shard; segments "pick up the marker", "write the letter W/U/J/I on the paper", "put down the marker" (successful segments of failed episodes included)	one flat marker lying at the far left, an A5 sheet (USD quad with a dynamic ink texture) in front of the left arm (centre (-0.10, -0.16) +- 2 cm, yaw +- 8 deg); thumb vs index+middle pinch at the marker's middle, rolled upright, each stroke servoed on the marker's tip; ink drawn where the tip touches; success = ink vs. template recall and precision >= 0.8 (6 mm tolerance); no arm noise. The host env is general_pickup (the sheet is not a layout object)

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