Overview of the Application of Deep Reinforcement Learning in Autonomous Driving Systems

Deep reinforcement learning (DRL) has both the processing ability of deep learning (DL) for high-dimensional input and the decision-making ability of reinforcement learning (RL), and can realize direct mapping from high-dimensional perception information to continuous action space output, which is very suitable for processing the autonomous driving tasks that complex environments and frequently interacting. The paper introduces the categories and research progress of DRL, analyzes the key technologies of autonomous driving system (ADS) in detail, discusses the application status of DRL in the key technical fields of environment perception, decision planning, and control execution of ADS, looks forward to the application prospects of DRL in ADS, and points out that research on the interpretability of DRL, improving the functional safety level, and research on the decision stability of DRL model, or research on use the DRL to improve the ability of comprehensive control in ADS have become the development direction.