JensenYuan/DexTacWAM_488_diverse_episodes

DexTacWAM — 488 diverse episodes

The tactile pretraining corpus for DexTacWAM. 488 episodes, 377,788 frames, 250 distinct task instructions, 3.5 hours at 30 fps on a Dexmate torso with two Sharpa Wave dexterous hands.

The corpus is built for coverage. It spans 250 instructions across 488 episodes — squeezing bubble wrap, cutting foil, shaking maracas, peeling adhesive wrap, extracting AirPods from a case — so most instructions appear only once or twice. It is intended for tactile representation learning; there are too few episodes per task to train a policy from it.

It is the only corpus used to train the released tactile encoder (464 episodes for training, the rest held out). The per-task demonstration datasets used for the world model and action experts are separate.

Format

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

feature	dtype	shape
tactile	uint8	(2, 5, 192, 256)	two hands x five fingers
tactile_flow	float32	(2, 5, 24, 32, 4)	per-taxel flow, precomputed
head_img	image	(192, 256, 3)	head camera, PNG-encoded
state	float32	(58)
actions	float32	(136)

Size

258 GB of episode data plus 12 GB of metadata (episodes_stats.jsonl is large because statistics are per-episode over high-dimensional tactile features). Episodes are independent files, so you can stream or pull a subset:

from huggingface_hub import hf_hub_download
p = hf_hub_download(
    "JensenYuan/DexTacWAM_488_diverse_episodes",
    "data/chunk-000/episode_000000.parquet",
    repo_type="dataset",
)

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}
}