Learning Playing Piano with Bionic-Constrained Diffusion Policy for Anthropomorphic Hand

Abstract

Anthropomorphic hand manipulation is a quintessential example of embodied intelligence in robotics, presenting a notable challenge due to its high degrees of freedom and complex inter-joint coupling. Though recent advancements in reinforcement learning (RL) have led to substantial progress in this field, existing methods often overlook the detailed structural properties of anthropomorphic hands. To address this, we propose a novel deep RL approach, Bionic-Constrained Diffusion Policy (Bio-CDP), which integrates knowledge of human hand control with a powerful diffusion policy representation. Our bionic constraint modifies the action space of anthropomorphic hand control, while the diffusion policy enhances the expressibility of the policy in high-dimensional continuous control tasks. Bio-CDP has been evaluated in the simulation environment, where it has shown superior performance and data efficiency compared to state-of-the-art RL approaches. Furthermore, our method is resilient to task complexity and robust in performance, making it a promising tool for advanced control in robotics.