Fusion of Reinforcement Learning and Simplified MPC for Path Tracking in Autonomous Vehicles

Model Predictive Control (MPC) determines control laws by predicting future system behavior and solving optimization problems, with wide applications in multiple fields. In autonomous vehicles, while MPC optimizes trajectories and controls, there are obvious problems. Parameter tuning is difficult. Frequent parameter testing is required under different road conditions and vehicle speeds, and tuning becomes more complex when the vehicle’s dynamic model deviates. Hard constraints are difficult to satisfy. Hard constraints such as steering angle and acceleration must be strictly observed. In complex environments, MPC cannot respond to emergencies in a timely manner, and it is easy to violate the constraints. This paper proposes the RL-SMPC framework: linearizing the vehicle dynamics state to a 3D state space, using Proximal Policy Optimization (PPO) to dynamically adjust MPC cost function weight matrices, and integrating Operator Splitting Quadratic Program (OSQP) for acceleration. MATLAB/Carsim simulations show that at 80km/h in double-lane-change conditions, compared with traditional MPC, lateral error is reduced by 42% (0.21m to 0.12m) and computational time by 60% (22.3ms to 8.5ms). It realizes online collaboration between policy gradient optimization and sparse QP solving, with Raspberry Pi 4B deployment verifying 10-ms level control frequency.