Diversity-Guided Genetic Algorithm for Safety-Critical Scenario Generation in Autonomous Driving Testing

Hua Qi, Siyuan Chen

Keywords: scenario-based testing, autonomous driving system, genetic algorithm

Abstract: As Autonomous Driving Systems (ADSs) rapidly advance and deploy in real-world settings, ensuring their safety and reliability has become paramount. While real-world testing provides comprehensive evaluation, it is prohibitively expensive and time-consuming. Simulation-based testing offers a practical alternative, with search-based methods demonstrating significant potential for efficiently identifying safety-critical scenarios. However, existing search-based approaches often prioritize efficiency over diversity, resulting in limited scenario coverage that may miss critical edge cases. To address this limitation, we propose a diversity-guided genetic algorithm for critical scenario generation. Our method introduces a diversity score to quantify scenario dissimilarity and computes the diversity rate between current scenarios and recent populations to guide the search process. We implement and evaluate our approach on the CARLA simulator across three distinct driving scenarios, comparing it against two baseline methods. Experimental results demonstrate that our method effectively and efficiently generates diverse safety-critical scenarios.

Submission Number: 15