Reproduce “CoorDex: Coordinating Body and Hand Priors for Continuous Dexterous Humanoid Loco-Manipulation” 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
Sikai Li, Shuning Li, Zhenyu Wei, Yunchao Yao, Chenran Li, Mingyu Ding
CoorDex facilitates continuous dexterous loco-manipulation for humanoids by coordinating separate latent priors for whole-body motion and multi-finger hand control within a residual reinforcement learning framework. This approach achieved successful dynamic grasping and object interaction while the robot was in continuous motion, outperforming direct joint-space control or monolithic latent policies on complex tasks.
Key insights
- Decomposing and independently training motion priors for the body and a wrist-stabilized hand significantly reduces the exploration complexity for high-DoF continuous loco-manipulation tasks.
- Utilizing latent action spaces for both body and hand control is crucial for trainability, as direct joint-space control and body-prior-only variants failed completely on complex dynamic tasks like WALKGRAB.
- A coordinated policy architecture with a shared task-level trunk and separate body/hand residual heads yields smoother whole-body motion and more reliable grasping compared to a monolithic latent residual policy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.