Crossing Traffic Avoidance of Automated Vehicle Through Bird-View Control, a Reinforcement Learning Approach

Abstract

This paper presents an innovative bird-view control framework for connected automated vehicles (CAV). Most recent tested automated vehicles are based on sensing systems equipped on the car body, which require the self-driving policy to be robust and adaptive to various environmental uncertainties. Inspired by the vehicle to infrastructure technologies, the self-driving technology can also be achieved through the communication between road infrastructure and the vehicle, where sensors are mainly installed on the road in a high position, which can collect traffic information from a bird-view. To this end, we developed a fusionbased Q-learning method to yield an optimal birdview control policy for a CAV on a single lane. With our control policy, the CAV can drive smartly under complicated traffic environment, interacting with leading vehicles and crossing traffic simultaneously. A series of case studies show our CAV control policy is string stable and can avoid collisions under various scenarios.