Robust biped locomotion using deep reinforcement learning on top of an analytical control approach

This paper proposes a modular framework to generate robust biped locomotion using tight coupling between an analytical walking approach and deep reinforcement learning. is composed of six main modules which are hierarchically connected reduce the overall complexity increase its flexibility. The core this specific dynamics model abstracts humanoid’s into two masses for modeling upper lower body. used design adaptive reference trajectories planner optimal controller fully parametric. Furthermore, learning developed based on Genetic Algorithm (GA) Proximal Policy Optimization (PPO) find optimum parameters learn how improve stability robot by moving arms changing center mass height. A set simulations performed validate performance official RoboCup 3D League simulation environment. results framework, not only in creating fast stable gait but also body efficiency.