Robust lane change decision for autonomous vehicles in mixed traffic: A safety-aware multi-agent adversarial reinforcement learning approach

Abstract

As autonomous driving technology advances, scenarios where autonomous vehicles coexist with human-driven vehicles will become increasingly common. However, existing lane-changing decision models often overlook the vulnerability to adversarial attacks and the complexity of traffic scenarios, leading to the risk of erroneous decisions in mixed traffic flows, which could result in catastrophic consequences. To address these issues, this paper proposes a novel dual-layer lane-changing decision-making algorithm to enhance the safety and reliability of autonomous vehicles in complex environments. The upper layer of the algorithm is based on a Multi-Agent framework, where adversarial learning is used to optimize decision strategies, allowing vehicles to adapt to various interferences and obstacles. The lower layer introduces a collision avoidance mechanism based on the principles of yielding and action masking, effectively reducing the risk of collisions. Experimental results under various traffic densities and disturbance conditions demonstrate that the proposed algorithm significantly improves the robustness of system, safety, and driving efficiency in complex conditions.