Reflex-Informed Neuromuscular Reinforcement Learning for Muscle-Driven Locomotion

Simulating muscle-driven human locomotion often struggles to balance physiological plausibility with adaptability to physical impairments. This paper introduces a reflex-informed neuromuscular reinforcement learning framework that anchors motion generation to a fixed phase-dependent reflex controller. Rather than directly solving complex muscle activations, the reinforcement learning policy modulates just four biomechanical residual parameters governing hip swing, knee support, and ankle propulsion. Under simulated muscle weakness and physical perturbations, the resulting locomotion maintains bilateral symmetry and stride consistency without policy retraining.