Adaptive MPC for Bipedal Robot Stable Walking Control

For bipedal robot locomotion control , the effectiveness of Model Predictive Control (MPC) in achieving stable Mass of Center (CoM) state tracking has been extensively validated. However, when bipedal robots traverse uneven terrain without relying on external sensors, the fixed prediction horizon and sampling frequency can hinder the advantages of MPC. This limitation leads to challenges in responding promptly to changes in the external environment and the state of bipedal robot, ultimately compromising safety and reducing resistance to uncertain disturbances. This paper presents a novel approach that integrates Proximal Policy Optimization (PPO) with MPC. In this framework, the PPO utilizes user control commands and tracking errors as input to dynamically output the state sampling frequency and prediction horizon. The sampling frequency is utilized to optimize the acquisition of input state data for the MPC, while the prediction horizon adjusts the dimensionality of the predictive equations and the associated Quadratic Programming (QP). By adaptively switching between sampling frequency and prediction horizon, this work significantly enhances the responsiveness of the MPC, enabling stable locomotion on uneven terrain and improving impact resistance.