Autonomous Vehicle Dynamic Following Based on Improved Reinforcement Learning

Target following technology, as one of the core technologies for autonomous vehicles, enables vehicles to automatically track dynamic targets. This paper proposes a path planning algorithm that combines proximal policy optimization(PPO) with Model predictive control(MPC) for dynamic target following tasks of autonomous vehicles in unknown environments. The algorithm utilizes neural networks to reduce computational load and employs MPC to compensate for the shortcomings of PPO in the prediction phase, thereby enhancing the following and obstacle avoidance capabilities of autonomous vehicles in dynamic environments. Simulation experiments show that the proposed PPO-MPC improves training speed by 35%, achieves trajectory following effects in low-speed dynamic target following tasks, and reduces the time required for static and dynamic obstacle avoidance tasks by approximately 8.5% compared to the original PPO, with smoother trajectories. It is capable of performing dynamic target following tasks in unknown environments with multiple obstacles.