🔥 News

• [2026.05.22] 🔥 We reprocessed a compliant embodied dataset on our university cluster. In this version, we removed non-compliant third-person source data and, after the full data pipeline, obtained a high-quality first-person robot video dataset with 3M videos, 720P resolution, and 30 FPS. To the best of our knowledge, RoVid-X is the largest robot video generation dataset to date for training physics-aware models. See Dataset | Resources.
• [2026.05.18] 🔥 We release StableVLA. Congratulations on its acceptance to ICML 2026! It is a vision-language-action model for robust robot policy learning. See Docs | Code | Project | Paper | Checkpoint.
• [Next Month] 🔥 We are preparing the open-source release of the HumanNet corpus, the curation pipeline, and the post-training validation code. Stay tuned!
• [2026.05.11]🔥 The HumanNet technical report and project page have been released: Paper | Project.

📑 Todo List

• Release the HumanNet technical report on arXiv. ✅
• Release StableVLA model code and documentation. ✅
• Release RoVid-X on Hugging Face. ✅
• Release a HumanNet preview subset on Hugging Face for early access.
• Release the full one-million-hour HumanNet corpus with metadata and annotations.
• Release the trained checkpoints initialized from HumanNet.

📣 Overview

This repository is maintained as a growing research hub for human-centric video data, embodied learning models, and validation code. It currently centers on HumanNet, a one-million-hour human-centric video corpus, and will also host related models, training recipes, evaluation protocols, and release notes.

The initial core release is HumanNet, a scalable infrastructure for fine-grained activity understanding, motion-aware video learning, and embodied pretraining. HumanNet pairs first-person and third-person footage with caption labels, motion annotations, and hand and body signals, organized by a multi-axis taxonomy and produced by a curation pipeline that treats human-centric filtering, viewpoint characterization, quality control, and privacy review as first-class design choices.

🎥 Demo

https://github.com/user-attachments/assets/52eaa410-0ec4-4f89-81e8-d2ecf9bb351c

📚 Dataset Family

🤖 Model Family

🗂️ Repository Map

HumanNet/
├── README.md                 # Repository entry point
├── docs/                     # Component-level documentation and release notes
│   ├── humandata.md          # HumanNet dataset documentation
│   └── stablevla.md          # StableVLA documentation
├── assets/                   # Figures used by the repository README
└── src/
    ├── dataset/
    │   ├── humandata/        # HumanNet dataset resources
    │   └── rovid-x/          # ROViD-X dataset resources
    └── model/
        └── StableVLA/        # StableVLA source code, training scripts, and model README

🔧 Usage

Coming soon.

# Download a HumanNet subset (placeholder)
# if you are in china mainland, run this first: export HF_ENDPOINT=https://hf-mirror.com
# pip install -U "huggingface_hub[cli]"
huggingface-cli download DAGroup-PKU/HumanNet

# Download RoVid-X
huggingface-cli download DAGroup-PKU/RoVid-X --repo-type dataset

🙏 Acknowledgement

We gratefully acknowledge SimpleSilicon Innovation for providing funding and resource support, and Astribot for providing real-robot platforms and deployment experiment support.

📧 Ethics Concerns

The videos referenced in this repository are sourced from public domains and intended solely to showcase the capabilities of this research. Human-centric video raises non-trivial privacy, consent, and dual-use concerns; any release will follow license review, redaction, restricted-content filtering, access controls where necessary, and clear documentation of what is included or excluded.

• The service is a research preview. Please contact us if you find any potential violations.

✏️ Citation

If you find our work useful in your research, please consider giving a star :star: and citation :pencil:.

BibTeX

@article{deng2026humannet,
  title={HumanNet: Scaling Human-centric Video Learning to One Million Hours},
  author={Deng, Yufan and Zhou, Daquan},
  journal={arXiv preprint arXiv:2605.06747},
  year={2026}
}

@misc{fu2026stablevlarobustvisionlanguageactionmodels,
      title={StableVLA: Towards Robust Vision-Language-Action Models without Extra Data}, 
      author={Yiyang Fu and Chubin Zhang and Shukai Gong and Yufan Deng and Kaiwei Sun and Qiyang Min and Qibin Hou and Yansong Tang and Jianan Wang and Daquan Zhou},
      year={2026},
      eprint={2605.18287},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.18287}, 
}