Deep Reinforcement Learning Based Tracking Control of Unmanned Vehicle with Safety Guarantee

Abstract

It is well known that the development of efficient real-time path following strategy and collision avoidance mechanism is critical to the practical implementation of autonomous driving technique. Within this context, this paper presents a new kind of hybrid control strategy consisting of the robot Stanley's trajectory tracking algorithm [1] and deep reinforcement learning (DRL) technique to achieve the goal of tracking control of unmanned vehicle with safety guarantee. By introducing the DRL technique, the tracking accuracy of the robot Stanley's trajectory tracking algorithm is improved and a safe control algorithm with collision avoidance is obtained. Furthermore, the complexity of the learning algorithm involved in the tracking controller is significantly reduced by using the Stanley's trajectory tracking algorithm, which makes the learning converge fast. Finally, numerical simulations are performed to verify that the proposed tracking algorithm has obviously advantages on tracking accuracy and training efficiency over some existing ones.