An LLM-enhanced Multi-objective Evolutionary Search for Autonomous Driving Test Scenario Generation

The safety of Autonomous Driving Systems (ADSs) is significantly important for the implementation of autonomous vehicles (AVs). Therefore, ADSs must be evaluated thoroughly before their release and deployment to the public. How to generate diverse safety-critical test scenarios is a key task for ADS testing. This paper proposes LEADE, an LLM-enhanced scenario generation approach for ADS testing, which adopts the LLM-enhanced adaptive evolutionary search to generate safety-critical and diverse test scenarios. LEADE leverages LLM's ability in program understanding to better comprehend the scenario generation task, which generates high-quality scenarios of the first generation. LEADE adopts an adaptive multi-objective genetic algorithm to search for diverse safety-critical scenarios. To guide the search away from the local optima, LEADE formulates the evolutionary search into a QA task, which leverages LLM's ability in quantitative reasoning to generate differential seed scenarios to break out of the local optimal solutions. We implement and evaluate LEADE on industrial-grade full-stack ADS platform, Baidu Apollo. Experimental results show that LEADE can effectively and efficiently generate safety-critical scenarios and expose 10 diverse safety violations of Apollo. It outperforms two state-of-the-art search-based ADS testing techniques by identifying 4 new types of safety-critical scenarios on the same roads.