A dual-stage reward mechanism for autonomous vehicle path-following via deep reinforcement learning

This paper investigates the path-following control of autonomous vehicle by proposing a dual-controllers deep deterministic policy gradient. Dual-controllers match with the lateral and longitudinal control of vehicle, respectively. The main contributions are threefold: (1) We consider six lateral and three longitudinal state variates separately, aiming to reduce lateral and longitudinal coupling. (2) A dual-controllers is designed correspondingly to control lateral steering angle and longitudinal acceleration. It utilizes a shared experience replay pool to achieve efficient training. (3) A novel dual-stage reward mechanism is designed for the two controllers, which combines direct error feedback and indirect incremental guidance to balance lateral tracking accuracy and longitudinal safety coordination. This design yields better training performance compared to a simple reward function. Finally, MATLAB is employed for experimental verification, while CarSim simulations are utilized for physical validation. The simulation results show that it outperforms other methods.