FlashDexRetarget

Accelerating Dexterous Manipulation Data Generation through Multi-Motion Retargeting

Kyungmin Lee^1*, Sibeen Kim^1*, Dongyoon Hwang^1*, Yoonsang Oh^1*, Donghu Kim², Youngdo Lee², I Made Aswin Nahrendra², Jaegul Choo^1†, Hojoon Lee^2†

¹KAIST AI   ²Holiday Robotics   ^*Equal contribution   ^†Corresponding author

Project page · arXiv · Paper · Video

FlashDexRetarget retargets a whole collection of human hand-object demonstrations to a dexterous robot hand with a single reference-conditioned RL policy: 90% success on a 50-motion benchmark with about 100x less training compute than CHORD.

Code is coming soon. The project page is served by GitHub Pages from docs/ on main.

Citation

@article{lee2026flashdexretarget,
  title   = {FlashDexRetarget: Accelerating Dexterous Manipulation Data Generation through Multi-Motion Retargeting},
  author  = {Lee, Kyungmin and Kim, Sibeen and Hwang, Dongyoon and Oh, Yoonsang and Kim, Donghu and
             Lee, Youngdo and Nahrendra, I Made Aswin and Choo, Jaegul and Lee, Hojoon},
  journal = {arXiv preprint arXiv:2610.01849},
  year    = {2026}
}