Safe Lane-Changing in CAVs Using External Safety Supervisors: A Review

Abstract Connected autonomous vehicles (CAVs) can exploit information received from other in addition to their sensor make decisions. For this reason, deployment is expected improve traffic safety and efficiency. Safe lane-changing a significant challenge for CAVs, particularly mixed traffic, i.e. with human-driven (HDVs) on the road, as set of around them varies very quickly, they only communicate fraction them. Many approaches have been proposed, most recent work adopting multi-agent reinforcement learning (MARL) approach, but those do not provide guarantees making unsuitable such safety-critical application. A number external techniques shielding, control barrier functions, model predictive recovery RL, applied CAV lane changing. This paper investigates whether supervisors could be used MARL-based changing (LC-CAV). purpose, MARL approach (MARL-CAV) designed, using parameter sharing replay buffer motivate cooperative behaviour collaboration among CAVs. then baseline discuss applicability state-of-the-art MARL-CAV. Comprehensive analysis shows that integrating an technique CAVs challenging, none existing directly MARL-CAV these require prior knowledge unsafe states policies.