UCLA

We present a biomechanics-based framework for the locomotion of a muscle-actuated human model that is driven by Central Pattern Generators (CPGs). Our CPG system directly generates the activation signals of 22 lower body muscles to reproduce the oscillatory patterns of locomotive bipedal stepping. We employ a dual-module architecture that trains a CPG Tuner network and a Reflex Controller to jointly adjust the muscle signals during simulation in order to adapt to the challenges of 3D bipedal locomotion. These modules are trained simultaneously using the Soft Actor-Critic (SAC) reinforcement learning algorithm. Our CPG system achieves stable, realistic walking and we observe promising results toward task-driven locomotion and adjustable gaits.