Decentralized Motor Skill Learning for Complex Robotic Systems

Reinforcement learning (RL) has achieved remarkable success in complex robotic systems (eg. quadruped locomotion). In previous works, the RL-based controller was typically implemented as a single neural network with concatenated observation input. However, corresponding learned policy is highly task-specific. Since all motors are controlled centralized way, out-of-distribution local observations can impact global through coupled policy. contrast, animals and humans control their limbs separately. Inspired by this biological phenomenon, we propose Decentralized motor skill (DEMOS) algorithm to automatically discover groups that be decoupled from each other while preserving essential connections then learn decentralized Our method improves robustness generalization of without sacrificing performance. Experiments on humanoid robots demonstrate robust against malfunctions transferred new tasks.