Reproduce “Dexterous Manipulation Transfer via Progressive Kinematic-Dynamic Alignment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenbin Bai, Qiyu Chen, Xiangbo Lin, Jianwen Li, Quancheng Li, Hejiang Pan, Yi Sun
The Progressive Kinematic-Dynamic Alignment (PKDA) framework translates human hand manipulation from demonstration videos into stable dexterous robot manipulation trajectories without extensive robot-specific data. This method achieved up to 77.5% transfer success rate across various robotic hands and tasks while demonstrating robustness to visual perception errors and successful real-world execution.
Key insights
- Deemphasizing fingertip-to-wrist vectors in kinematic mapping and focusing on absolute fingertip positions improves precision and reduces sensitivity to morphological differences.
- Constraining RL exploration through thumb-guided pre-grasp initialization and action space rescaling significantly boosts learning efficiency and prevents unstable behaviors.
- A hierarchical unified reward design, combining contact, imitation, and lifting rewards, effectively guides the robot towards stable multi-point grasps and task completion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.