A Novel Reinforcement Learning Method for Autonomous Driving With Intermittent Vehicle-to-Everything (V2X) Communications
Longquan Chen, Ying He, F. Richard Yu, Weike Pan, Zhong Ming
Abstract
Most autonomous driving studies incorporating vehicle-to-everything (V2X) communications assume the existence of continuous V2X communications, and ignore the possibility that V2X communications can be interrupted intermittently. However, when V2X communications are interrupted, autonomous vehicles that rely on the information from V2X communications can fall into disasters. To address these issues, we propose a novel reinforcement learning method named RL4V2X for decision-making and motion-controlling of autonomous driving. This method is composed of a convolutional neural network (CNN), a gate recurrent unit (GRU) and three gate networks. Specifically, the CNN extracts a vehicle spatial distribution to better represent the traffic. The GRU constructs the sequence features to characterize the dynamics of traffic, which can be used to compensate for the lack of information when V2X communications are interrupted. In addition, the gate networks assign the confidence to different features according to the interruption situations. Extensive simulation results in different traffic conditions demonstrate the superior performance of the proposed method.