Adaptive safety performance testing for autonomous vehicles with adaptive importance sampling

Efficient and accurate safety testing and evaluation are crucial for autonomous vehicles (AVs). Recent studies have utilized prior information, such as surrogate models, to enhance testing efficiency by deliberately generating safety-critical scenarios. However, discrepancies between this prior knowledge and actual AV performance can undermine their effectiveness. To address this challenge, adaptive testing methods dynamically adjust testing policies based on posterior information of AVs, such as testing results. Most existing approaches focus on adaptively optimizing testing policies during pre-tests, yet neglecting how to adapt the testing policies in the large-scale testing process that is required for unbiased safety performance evaluation. To fill this gap, we propose an adaptive testing framework that continuously optimizes testing policies throughout large-scale testing. Our approach iteratively learns AV dynamics through deep learning and optimizes testing policies based on the learned dynamics using reinforcement learning. To tackle the challenge posed by the rarity of safety-critical events, our method focuses exclusively on learning safety-critical states in both the dynamics learning and the policy optimization processes. Additionally, we enhance evaluation robustness by integrating multiple pre-trained testing policies and optimizing their combination coefficients. To accurately assess safety performance, we evaluate testing results obtained from various testing policies using adaptive importance sampling. Experimental validation in overtaking and unprotected left-turn scenarios demonstrates the significant evaluation efficiency of our method.