Bipedal Robot Gait Control Strategies Based on Reinforcement Learning

This paper presents a reinforcement learning-based gait control strategy for bipedal robots. The work focuses on enabling the robot to adapt to complex terrains without relying on visual information. By leveraging the neural network's fitting capability in a simulated environment, the robot estimates privileged information from limited proprioceptive data. The approach is based on reinforcement learning fundamentals and includes a teacher-student policy architecture. We design a series of reward functions to achieve velocity tracking, periodic gait phases, accurate foot traj ectories, and smooth locomotion. Domain randomization is employed to bridge the gap between simulation and reality. Experiments using Proximal Policy Optimization show promising results, and the future work aims to deploy the policy on a real humanoid robot and explore full-body walking strategies.