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 scaling robot learning.

Status: Upload complete · 26.67 TB, 694,310 payload files uploaded and verified · Upload details.

Open-HumanWAM-50K accompanies Open-HumanWAM: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models. The corpus combines HumanNet, EgoSuite-Open100K, EgoDex, Gen-HumanEgo, and EgoVerse, spanning household, commercial-service, and industrial manipulation.

The data 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 report describes 35.9 million selected clips; this release packages 31,396,799 physical episodes. Selected clips and stored episodes are different counting units.

Dataset construction: multi-source egocentric video with language and hand-motion annotations, visibility and motion-quality filtering, and alignment of camera frames, wrist axes, and gripper openness.

✅ DATA UPLOAD COMPLETE · 数据上传完成

All dataset payload files have been uploaded and verified. This includes the base corpus and supplemental data for the complete historical subset selections.

Upload status	Final result
Uploaded and verified	26.67 TB / 26.67 TB (100%)
Payload files	694,310 / 694,310
Completed	2026-11-11 07:38 (UTC+8)
Last updated	2026-11-11 12:56 (UTC+8)

Payload totals include video, action/state data, metadata, and provenance files from the base corpus and supplement. Each uploaded payload file was verified against its expected size and SHA-256. TB uses decimal units.

Format and organization

The release uses LeRobot v3, with compressed MP4 videos, Parquet action/state data, captions, episode metadata, and source provenance. Text embedding caches are excluded; captions are retained.

egostandard/part-*/
egopro/part-*/
egodex/part-*/
egoverse/part-*/
humannet/part-*/

Each part-* is an independent LeRobot v3 dataset with data/, videos/, and meta/. EgoStandard and EgoPro come from EgoSuite-Open100K; EgoVerse uses the selected Mecka subset. The humannet/ collection also contains Gen-HumanEgo and other retained sources; source identity is preserved in provenance.

Training subset indices

Choose a training subset · Coverage and checksums

Each subset is a self-contained Parquet table of dataset_path, episode_index, start_frame, end_frame, source, task, and hand. These addresses point directly into the released LeRobot parts. No historical root IDs or provenance joins are needed. Named Hugging Face configurations load these index rows; video/action payloads are downloaded separately.

All seven subset indices (100h, 500h, 1000h, 5000h, 10000h, 20500h and 50000h) cover their complete historical selections. Additional egostandard/part-supp1105-* payloads supply the formerly missing small-subset intervals and are uploaded and verified. The 50,000h selection is unchanged. The base corpus payload upload is complete and verified. No replacement clips or validation holdouts are introduced.

See collection.json for collection-level metadata. Source-specific dataset terms apply; this collection does not replace the original source licenses.

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}
}