A cooperative lateral and vertical control strategy for autonomous vehicles based on multi-agent deep reinforcement learning

With the increasing level of automation in autonomous vehicles, consideration of comfort and stability will further enhance the public acceptance of autonomous driving technology. This paper presents a cooperative lateral and vertical control strategy for autonomous vehicles based on multi-agent deep reinforcement learning, which integrates path tracking and suspension control for different planar learning tasks. By developing the lateral and vertical dynamic models, the multi-objective coordinated exploration of path tracking and active suspension systems is imposed by using the deep deterministic policy gradient (DDPG) algorithm. In the multi-agent deep reinforcement learning, a feedforward steering of steering subsystem and a PID compensation control of suspension subsystem are added to the DDPG control process for efficiently searching the strategic action of the coupling system. Furthermore, the learning reward function of autonomous vehicle is designed by comprehensively considering the accuracy, safety and comfort performance. Through the trained learning process and simulation results under different driving conditions, the proposed method can achieve the simultaneous optimization of path tracking and suspension comfort performance, and effectively improve the ride comfort and stability in the high-performance path tracking process. This study provides an efficient control scheme for improving the ride comfort of autonomous vehicles.