Prior-Knowledge-Guided Reinforcement Learning for Musculoskeletal Locomotion

Abstract

Reinforcement learning (RL) has achieved notable success in locomotion of robotic systems, yet remains challenged by high-dimensional and overactuated musculoskeletal systems. In this work, we propose a synergy-guided RL framework that incorporates biologically inspired priors extracted from pre-trained locomotion policies. Specifically, we derive lowdimensional muscle activation representations and functional muscle group structures from coordinated activation patterns. These priors are embedded into policy learning, enabling the control of structured muscle synergies instead of individual muscles. By reducing effective action dimensionality while preserving intrinsic coordination patterns, the proposed approach improves sample efficiency and training stability. Extensive experiments demonstrate that our method achieves fast convergence, high rewards, and robust locomotion under standard training conditions, while also generalizing effectively to unseen environmental settings, including stair, hilly, and rough terrain.