Hybrid Control of Advanced Quadruped Locomotion: Integrating Model Predictive Control with Deep Reinforcement Learning

This paper presents a comprehensive framework for controlling a quadruped robot using a combination of Model Predictive Control (MPC) and Reinforcement Learning (RL). The objective is to improve the robot's locomotion efficiency and stability in dynamic environments. We model the system using a state-space approach, derive the robot's kinematics, and implement Proximal Policy Optimization (PPO) for RL training. Simulations are conducted for both trot and gallop gaits, where the robot's performance is compared using pure MPC and hybrid RL+MPC (RLMPC) control strategies. The results demonstrate that the RLMPC method improves velocity tracking and allows for faster completion of tasks compared to pure MPC, particularly in dynamic settings. The simulations also highlight areas where control smoothness could be improved, especially in reducing chattering during motion. Training and computational performance are evaluated, emphasizing the trade-off between parallelism and efficiency. The findings suggest that combining learning-based approaches with predictive control enhances the quadruped's ability to adapt to complex terrains and tasks.