Reproduce “Learning Diverse Bimanual Dexterous Manipulation Skills from Human Demonstrations” on OpenResearch

About this paper

Bohan Zhou, Haoqi Yuan, Yuhui Fu, Zongqing Lu

Researchers from Peking University developed BiDexHD, a framework that constructs diverse bimanual dexterous manipulation tasks from human demonstrations and learns policies via a teacher-student reinforcement learning paradigm. The system demonstrated high success rates on trained tasks and competitive zero-shot generalization to tasks involving previously unseen objects, addressing the need for scalable bimanual skill acquisition.

Key insights

• Independent learning for each hand (IPPO) effectively acquires robust bimanual skills and adapts better to diverse tasks compared to a centralized PPO approach.
• An explicit two-stage reward function with an initial alignment phase is critical for successful bimanual manipulation, particularly when combined with functional grasping centers derived from human demonstrations.
• Distilling state-based teacher policies into a single vision-based student policy, conditioned on visual features like point clouds, enables competitive zero-shot generalization to tasks involving unseen objects.

Summary by alphaXiv.