Lane Keeping Assist for an Autonomous Vehicle Based on Deep Reinforcement Learning 2020-01-0728
Lane keeping assist technique plays an important role in autonomous driving as it keeps the vehicle travelling along a desired line of the lanes by adjusting the front steering angle. Reinforcement learning (RL) can be used to teach machines through interaction with the environment and learn from their mistakes. The model-free characteristics of reinforcement learning free us from coding complex policies manually. But it has not yet been successfully used for automotive applications. Recently, deep reinforcement learning (DRL) which integrates reinforcement learning and deep neural network (DNN) has achieved progress greatly in domains such as learning to play Atari games from raw pixel input. In this paper, a control strategy using two different state-of-the-art deep reinforcement learning algorithms has been proposed and used in the lane keeping assist scenario. Deep Q-Network Algorithm (DQN) with discrete action space and Deep Deterministic Policy Gradient Algorithm (DDPG) with continuous action space have been implemented, respectively. Based on the reinforcement learning toolbox and deep learning toolbox in Matlab, we design a deep neural network to represent the control policy and make decisions. The environment as well as the vehicle dynamics are modelled in Simulink. By integrating the proposed control method and a vehicle dynamics model, the lane keeping assist simulation was performed. Our results demonstrate that the vehicle travel along the centerline of the path and the controller reaches a steady state after a short time, implying the effectiveness of the proposed control method. Finally, the evaluation of results is given from some aspects of control efficiency and safety.
Qun Wang, Weichao Zhuang, Liangmo Wang, Fei Ju
Nanjing University of Science and Technology, Southeast University