Design and Experimental Validation of Lane Centering Control and Dynamic Obstacle Avoidance for Uncrewed Autonomous Car

Accurate lane centering control (LCC) and dynamic obstacle avoidance (DOA) play a vital role in guaranteeing the safety and stability performance of uncrewed autonomous cars (UAC) in unknown dynamic transportation environments. Conventional reinforcement learning algorithms used for the LCC and DOA of the UAC may suffer from defects, such as data insufficiency and training inefficiency. To this end, an innovative control algorithm that integrates dynamic information into a deep deterministic policy gradient (DDPG) controller, referred to as DI-DDPG for UAC, is proposed. First, a dynamic information-integrated feature extraction scheme is developed to enhance the training efficiency and convergence rate. Then, a custom self-decaying noise is incorporated into the actions output to improve the convergence rate of the algorithm. Moreover, a black frame image clipping technology is proposed to enhance the accuracy of action generation. Finally, comparative simulations are carried out in a Racetrack environment, and then field experiments are conducted to demonstrate the practical effectiveness of the proposed DI-DDPG controller.