Joint Local Reinforcement Learning Agent and Global Drone Cooperation for Collision-Free Lane Change

This chapter introduces a drone-assisted lane change platform with joint local and global control for collision-less lane change. Specifically, the local control is based on a reinforcement learning agent, DEAR (DEep Q-network with a dynAmic Reward), while the global control is based on drones. The reward function of DEAR is designed from safety, comfort, and efficiency perspectives, and the weights of the three rewards are adjusted according to the surrounding traffic condition. On the other hand, the drones hovering over the highway provide global information (i.e., road vehicular density) to the ego vehicle while performing global control by: (1) computing and sending a dynamic collision reward to the ego vehicle; (2) sending an urgent lane change request (ULCR) to the ego vehicle when a road risk ahead, or an emergency vehicle behind the ego vehicle is detected. The proposed lane change platform is tested with the authentic next-generation simulation (NGSIM) dataset. Simulation results prove that the platform is able to perform safe and efficient lane change on a road prone to risks and emergency vehicles.