Abstract
Redundant muscles in human-like musculoskeletal robots provide additional dimensions to the solution space. Consequently, the computation of muscle excitations remains an open question. Conventional methods like dynamic optimization and reinforcement learning usually have high computational costs or unstable learning processes when applied to a complex musculoskeletal system. Inspired by human learning, we propose a phased target learning framework that provides different targets to learners at varying levels, to guide their training process and to avoid local optima. By introducing an extra layer of neurons reflecting a preference, we improve the Q-network method to generate continuous excitations. In addition, based on information transmission in the human nervous system, two kinds of biological noise sources are introduced into our framework to enhance exploration over the solution space. Tracking experiments based on a simplified musculoskeletal arm model indicate that under guidance of phased targets, the proposed framework prevents divergence of excitations, thus stabilizing training. Moreover, the enhanced exploration of solutions results in smaller motion errors. The phased target learning framework can be expanded for general-purpose reinforcement learning, and it provides a preliminary interpretation for modeling the mechanisms of human motion learning.