Path tracking control based on Deep reinforcement learning in Autonomous driving

Lane keep assist (LKA) and Adaptive Cruise Control (ACC) are two fundamental yet critical functions for autonomous driving, and conventional methods using PID controllers may not perform well in certain extreme driving conditions. In this paper, we propose a reinforcement learning based approach to train the agent to learn LKA and ACC and hence adapt to diverse scenarios. Particularly, we employ deep deterministic policy gradient (DDPG) algorithm to train the agent and consider both state space and action space as continuous, and designed two neural network critic-network and actor-network to simulate the strategy function and Q-function. Then, we train the two neural networks by deep learning method. Finally, Simulations are conducted with both reinforcement learning and traditional PID controller, and the results of reinforcement learning is more adaptive to extreme road conditions in comparison with a traditional PID controller.