A Distributed Driving Decision Scheme Based on Reinforcement Learning for Autonomous Driving Vehicles

Autonomous driving technology is one of the research hotspots in recent years. The development of some basic technologies such as motor sensors, vehicle to everything (V2X) and high-definition map also provides more powerful technical support for autonomous driving. However, it remains a challenge to propose a common decision strategy for autonomous driving. Considering the scenario that the networks can cover various urban roads through road side units (RSUs), this paper proposes a two-satage autonomous driving decision strategy based on environmental awareness and asynchronous advantage actor critic (A3C) algorithm. The first stage proposes a preliminary common model with safety guarantee for all roads; based on which the second stage trains a unique model for each road in consideration of different road features. Simulation results indicate that the proposed strategy can ensure high safety performance and driving efficiency.