DAGroup-PKU/Open-HumanWAM-50K

Open-HumanWAM

An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models

Report · Code · Project page · Models · HF Paper · Citation

Open-HumanWAM overview: scaling human and robot experience for embodied foundation models.

Open-HumanWAM-50K

50,000 hours of egocentric human manipulation video with hand-motion and natural-language annotations for robot learning.

Open-HumanWAM-50K is the human-data corpus accompanying Open-HumanWAM: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models. It brings together HumanNet, EgoSuite-Open100K, EgoDex, Gen-HumanEgo, and EgoVerse, covering household, commercial-service, and industrial manipulation.

The corpus is curated for hand visibility, motion stability, and manipulation activity. Cross-source alignment standardizes camera coordinates, wrist axes, and hand/gripper openness into a unified bimanual end-effector representation. The 50,000-hour selection comprises 35.9 million selected clips drawn from 31,396,799 physical episodes; multiple clips can reference the same episode.

From diverse egocentric recordings to curated video, language, and hand-motion data for robot learning.

Format and organization

Data is stored in LeRobot v3 format: compressed MP4 video, Parquet action/state data, captions, episode metadata, and source provenance. Captions are included; precomputed text embeddings are not.

Directory	Source grouping
egostandard/	EgoStandard from EgoSuite-Open100K
egopro/	EgoPro from EgoSuite-Open100K
egodex/	EgoDex
egoverse/	EgoVerse, corresponding to the selected Mecka data
humannet/	HumanNet, Gen-HumanEgo, and other retained sources

Each directory contains part-* folders. Each part is an independent LeRobot v3 dataset with data/, videos/, and meta/; provenance/ preserves the original source identity. See collection.json for collection metadata.

Choose a training subset

Seven subset configurations support experiments at different scales: 100h, 500h, 1000h, 5000h, 10000h, 20500h, and 50000h. Each preserves the corresponding historical training selection, including the supplemental episodes referenced by the smaller subsets.

Subset guide · Subset manifest

Each subsets/<size>.parquet file is a self-contained clip index with these fields:

Field	Meaning
dataset_path	Relative path to a LeRobot part
episode_index	Episode identifier within that part
start_frame, end_frame	Episode-local frame interval, with an exclusive end
source	Original data source
task	Natural-language annotation for the clip
hand	Hand annotation used by the training selection

The Hugging Face configurations load these clip indices. Download the referenced LeRobot parts separately to access the video and action/state data.

Access and training

To access the data, sign in to Hugging Face and complete the access request form. Requests are approved automatically. For scripted downloads, authenticate with a token from the approved account, for example with hf auth login.

For model setup and training instructions, see the Open-HumanWAM code repository.

License

This project's original contributions are released under the Apache License 2.0. The underlying third-party datasets remain subject to their original licenses and terms. This release does not replace or expand the permissions granted by those source datasets; applicable source terms continue to apply to their content.

Citation

If you use Open-HumanWAM-50K in your research, please cite our technical report:

@misc{ma2026openhumanwam,
  title = {{Open-HumanWAM}: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models},
  author = {Ma, Juncheng and Li, Jiaxin and Tu, Jiankai and others},
  year = {2026},
  eprint = {2611.06817},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2611.06817}
}