DexSeed

Expanding Human Hand–Object Interactions into Robot Demonstrations for Dexterous Policy Learning

Huaijin Pi^1,2, Zehong Shen^2,*, Yong Wang³, Kechun Xu², Changpeng Wang³,
Dingkun Wei^2,3, Xueyang Wang^2,3, Wenjia Wang¹, Xing Zhu², Yujun Shen^2,†, Taku Komura^1,†

¹ The University of Hong Kong    ² Robbyant    ³ Zhejiang University
^* Project Lead    ^† Corresponding Authors

Project Page & Demos  ·  arXiv: Coming soon

From human interactions to real-world dexterous manipulation. DexSeed expands human hand–object interactions into physically executable robot demonstrations across object geometries and manipulation locations. A VLA fine-tuned solely on the generated demonstrations performs language-specified grasps on a real robot, without collecting real-robot action demonstrations.

Code release

Coming soon. We are preparing the code release. The implementation and instructions for using it will be added to this main branch.

Explore the method, videos, and experimental results on the project page. Its source and media are maintained on the gh-pages branch.

Citation

If you find DexSeed useful for your research, please cite:

@misc{pi2026dexseed,
  title = {DexSeed: Expanding Human Hand--Object Interactions into Robot Demonstrations for Dexterous Policy Learning},
  author = {Huaijin Pi and Zehong Shen and Yong Wang and Kechun Xu and Changpeng Wang and Dingkun Wei and Xueyang Wang and Wenjia Wang and Xing Zhu and Yujun Shen and Taku Komura},
  year = {2026},
  url = {https://robbyant-research.github.io/DexSeed/}
}

_{Huaijin Pi, Yong Wang, Changpeng Wang, Dingkun Wei, and Xueyang Wang conducted this work during their internships at Ant Group.}