上海科技大学知识管理系统

Autonomous Driving Systems (ADS) are inherently safety-critical, necessitating comprehensive evaluation, especially within safety-critical scenarios. Ensuring that evaluation scenarios are both authentic and interpretable is essential for developers to effectively analyze and address identified issues. Current ADS evaluations predominantly rely on road testing supplemented by open-loop simulations with predefined Non-Player Character (NPC) trajectories, which fail to capture complex interactions and lack scenario diversity. To address these limitations, we introduce the Adversarial Cognitive Driver Model (ACDM), a mechanism-based driver behavior model designed for closed-loop testing. ACDM employs a utilitarian framework to calculate expected rewards and risks when determining control actions, prioritizing its own safety before introducing adversarial rewards that incorporate relevance and risk factors towards the ADS. Additionally, we propose a model-based initial condition sampling method that leverages the internal cognitive states of ACDM to generate a diverse array of scenarios independent of physical states. Implemented within the open-source CARLA framework, ACDM facilitates comprehensive and accessible ADS testing. We conducted experiments using two driver behavior models as ADS surrogates and compared ACDM against the state-of-the-art adversarial driver behavior model, Dense Deep Reinforcement Learning (D2RL). The evaluation focused on the relevance and safety-criticalness of various experimental setups in both short-term and long-term simulations. Results demonstrate that ACDM outperforms its counterparts in maintaining relevance towards the ADS, particularly in longterm simulations. Furthermore, ACDM generates diverse, interpretable, and rational driver behaviors, enhancing the authenticity of interactions. Our model-based scenario generation method based on internal cognitive states also significantly improves the diversity of closed-loop simulations. These advancements provide a robust framework for evaluating and enhancing the safety of ADS in dynamic driving environments.