Cyan666/HandsOn-Datasets

HandsOn Demonstrations

Project homepage · Paper (arXiv) · Code

HandsOn provides 18,000 curated simulated dexterous-manipulation demonstrations in LeRobot v2.1 format. Each episode contains measured robot states, commanded actions, language instructions, and synchronized RGB from three cameras. The data supports training and offline analysis for the HandsOn benchmark. Closed-loop success is measured with the simulator and task checkers in the code repository; action prediction error is not a success rate.

Dataset summary

Property	Value
Episodes	18,000
Frames	6,053,555 at 30 Hz (about 56.1 hours)
Training split	15,178 episodes from 895 collection sessions
Validation split	2,822 episodes from 155 collection sessions
Videos	54,000 (three per episode)
Cameras	Front: 854 × 480; left/right wrist: 848 × 480
State / action	52 dimensions each; right side first, then left
Collection platform	ROKAE AR5-5_08 arms and Wuji Hand 1

The split separates collection sessions; no episode or collection session occurs in both splits. Normalization statistics use the training split only. Validation is held out for training that respects this split; it is not necessarily unseen by models previously trained on the source recordings.

Task scope

The benchmark contains 50 Core tasks and 16 evaluation-only Transfer tasks. The released demonstration corpus has 54 native source identities organized into 52 selection quota groups and 91 active task/prompt pairs. These are different counting units: collection stages retain their own source identities, and the corpus retains chess recordings excluded from the benchmark evaluation. The total_tasks value of 120 in meta/info.json counts the language/task lookup dictionary, including entries not active in this subset; it is not a claim of 120 benchmark tasks. See the paper for the task inventory and scope.

The first preview episodes prioritize tasks implemented for HandsOn. Preview ordering does not change the split, source trajectories, or benchmark membership.

Features and action convention

observation.state is the measured 52-dimensional state. action is the aligned absolute world-palm command, not a substituted future measured state. Each side occupies 26 dimensions: three palm-position values, three palm rotation values, and 20 hand-joint values. The right side precedes the left. Use the release's state/action convention and normalization together with the simulator interface; raw action values must not be treated as joint velocities.

Images use the released robot, camera, room, table and lighting setup at 32 samples per pixel. Front and wrist images have the resolutions above. Native task resets, recorded numerical values, source identities and episode seeds are retained. Episode and global-frame indices are contiguous within this release. Historical provenance records retain their original identifiers.

Download and read

A full download is approximately 111 GB. Install the reader dependencies and retrieve the repository:

python -m pip install huggingface_hub numpy pyarrow av Pillow
hf download Cyan666/HandsOn-Datasets --repo-type dataset --local-dir ./handson-data

from pathlib import Path
import sys

root = Path("handson-data").resolve()
sys.path.insert(0, str(root / "code"))
from training_reader import HandsOnDataset

dataset = HandsOnDataset(root, horizon=16, split="train")
dataset.set_epoch(0)
sample = dataset[0]
print(sample["observation.state"].shape)  # (52,)
print(sample["action"].shape)             # (16, 52)
print(sample["action_is_pad"].shape)      # (16,)
print(sample["source_task"])

The reader indexes frames, not episodes. It returns the measured state, three RGB images, a future command chunk, an episode-end padding mask, an active-side dimension mask, and source task/prompt metadata. Terminal chunks repeat the final command; mask out entries where action_is_pad is true. decode_images=False supports numerical-only analysis. Set the epoch before constructing data-loader worker replicas for repeatable language variation.

Canonical prompts and 32 training / eight unseen paraphrases per base prompt are included. The default training sampler uses the canonical prompt with probability 0.2 and a training paraphrase with probability 0.8. Validation uses canonical prompts by default; use prompt_split="unseen" explicitly for held-out language evaluation. Episode split and prompt split are separate.

Files and provenance

• data/: per-episode Parquet state/action tables.
• videos/: synchronized front and wrist MP4 recordings.
• meta/: feature schema, episode/task dictionaries and split definitions.
• norm_stats.json: training-split normalization statistics.
• language/: canonical prompts and split-specific paraphrases.
• code/: standalone dataset reader and prompt sampler.
• source_manifest.jsonl: source identity, split and per-episode checksums.
• FILE_MANIFEST.jsonl: file sizes and SHA256 checksums.
• provenance/ and PACKAGE_COMPLETE.json: collection and publication records.

Source provenance is retained verbatim. Low brightness alone is not a corrupt recording; decoding, frame counts, finite numerical values and source checksums are the relevant integrity checks. This simulated corpus does not establish real-robot safety or replace closed-loop task evaluation.

License

The dataset and reader code are released under the MIT License. Please retain the copyright and license notices when redistributing them. Third-party source assets retain their respective licenses; the code repository notices document their attribution.

Citation

@article{luo2026handson,
  title = {{HandsOn}: A Large-Scale Simulation Benchmark for Generalizable Dexterous Manipulation},
  author = {Luo, Tianwei and He, Yiming and Wu, Hongjin and Pan, Ting and Ren, Debin and Guo, Haorui and Yang, Junlin and Zuo, Yuxin and Kou, Zhehao and Yang, Kaisen and Lai, Zhuoqun and Ding, Yi and Yu, Jiale and Zhang, Yuchen and Qiu, Zijie and Li, Haozhan and Ding, Ning},
  year = {2026},
  journal = {arXiv preprint arXiv:2610.26765},
  doi = {10.48550/arXiv.2610.26765},
  eprint = {2610.26765},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2610.26765}
}