KAIST AI and Holiday Robotics · Dexterous Hand Retargeting

FlashDexRetarget

Turns large collections of recorded human hand-object demonstrations into motions a dexterous robot hand can perform, training one reinforcement learning policy across all motions instead of one per demonstration. It reaches 90% retargeting success on a 50-motion XHand benchmark in 29 GPU-hours, which the authors put at about 100 times less compute than CHORD's 2,847, and converts more than 600 of 1,000 motions within 60 GPU-hours, author-reported. Tested on XHand and Sharpa Wave Hand on Isaac Sim 5.1 and Isaac Lab 2.3, with TACO, OakInk2 and HOT3D data pipelines. Paper arXiv 2610.01849 (October), MIT code.