Risk-guided efficient scenario selection for autonomous driving testing

Ensuring the Safety of the Intended Functionality (SOTIF) of autonomous driving systems (ADSs) remains a significant challenge, requiring comprehensive testing across diverse and complex environments. While existing works have contributed to scenario-based testing, their effectiveness is hindered by inadequate seed scenario selection and limited prioritization of high-risk scenarios, leading to inefficient identification of critical but rare traffic misbehaviors. To address these limitations, we propose EF-Fuzz, a novel testing framework designed to improve the efficiency of discovering safety-critical scenarios. EF-Fuzz incorporates two key modules: a risk-guided scenario risk prediction module and a seed selection and scheduling module. The risk prediction module utilizes various features from tested scenarios to accurately train a model that predicts the probability of safety risks in new scenarios. The seed selection and scheduling module enhances the fuzzing process by prioritizing high-risk seeds in the queue, thereby improving the efficiency of identifying problematic scenarios. Additionally, it balances scenario diversity and efficiency, avoiding local optima. In extensive evaluations using the Apollo autonomous driving system within the CARLA simulator, EF-Fuzz demonstrated a 37.9% increase in efficiency compared to state-of-the-art methodologies. It successfully identified 12 distinct types of corner cases, 10 of which were newly discovered. These results emphasize EF-Fuzz’s potential in advancing the safety testing of ADSs.