JensenYuan/DexTacWAM_unstack_bowl

Unstack Bowl by Pinching

One of the six evaluation tasks in DexTacWAM. 100 episodes, 36,712 frames at 30 fps, on a Dexmate torso with a Sharpa Wave dexterous hand (right only).

Hold the lower bowls with the right thumb, lift the edge of the top bowl with the index or middle finger, slide the thumb over, then pinch and lift the top bowl.

Separating the top bowl from a nested stack. The thumb holds the lower bowls while the index or middle finger lifts the edge of the top one, after which the thumb slides over to pinch it. The thumb can only slide in once that edge is actually free, and that moment registers in contact before it is visible from the cameras.

Format

LeRobot v2.1. Parquet under data/chunk-000/, metadata under meta/. 17 GB.

feature	dtype	shape
head_img	image	(192, 256, 3)	PNG-encoded
right_wrist_img	image	(192, 256, 3)	PNG-encoded
tactile	uint8	(1, 5, 192, 256)	one hand x five fingers
state	float32	(45)
actions	float32	(75)
tactile_flow	float32	(1, 5, 24, 32, 4)	per-taxel flow, precomputed
deform	uint8	(1, 5, 192, 256)	deformation rendering

episode_provenance.json records, per output episode, which raw episode and frame range it came from and why it was split there.

Intended use

Stage 2 (world model) and stage 3 (action expert) training. The configs live under configs/bowl_unstack/ in the code repository, and the normalization statistics committed there are computed against exactly this data. Statistics from a different conversion will de-normalize actions incorrectly without raising an error.

The configs address this corpus by directory name, so unpack it as data/datasets_lerobot/20260725_pinch_from_bowl_with_fingers_right_only/ and keep the published episode ordering; the validation split is selected by episode index.

The released tactile encoder was trained on the separate 488 diverse episodes corpus, not on this data.

We do not release a trained action expert for this dataset. In our experiments, the stage 3 action expert is randomly initialized and trained from scratch using this data.

License

Apache 2.0, matching the DexTacWAM code. Parts of that repository are additionally CC BY-NC-SA 4.0 where they derive from Genie-Envisioner; that restriction applies to those source files, not to this data.

Citation

@article{dextacwam2026,
  title   = {DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation},
  author  = {Yuan, Haoran and Wang, Zekai and Shao, Boning and Lu, Haoran and
             Darrell, Trevor and Lourentzou, Ismini and Zhan, Wei},
  journal = {arXiv preprint arXiv:2609.24976},
  year    = {2026}
}