Learning to Be Adversarial Yet Compliant: Safety-Critical Driving Scenario Generation via Safe Reinforcement Learning

Safety-critical scenario generation is a crucial technique for evaluating the reliability of autonomous vehicles (AVs) in infrequent but high-risk situations. However, existing methods often generate scenarios that are either excessively adversarial, leading to unrealistic results or traffic rule violations, or overly compliant, restricting their effectiveness in exposing critical vulnerabilities in AV systems. To address this problem, this paper proposes a novel hierarchical framework, Compliant Adversarial Scenario Generation (CASG), to enable the creation of scenarios that are both adversarial and compliant with traffic laws. Specifically, CASG employs a hierarchical safe deep reinforcement learning architecture with an integration of hard traffic rule constraints and adversariality. In this architecture, an action space pre-filtering mechanism is developed to ensure adherence to traffic rules and a hybrid advantage function is presented to optimize the balance between adversarial intensity and compliance. Performance investigation on the CARLA simulator and benchmark scenarios demonstrates that CASG achieves noticeable improvements in generating scenarios that are simultaneously highly adversarial and compliant, thereby enhancing the efficiency of AVs safety evaluation. The code and model will be available at https://github.com/BUCT-IUSRC/Research__CASG.