Large Language Model and Behaviour Tree Based Real-World Test Scenario Generation for Autonomous Vehicles

As autonomous driving systems (ADSs) move from experimental setups to real-world applications, ensuring safety and reliability remains a significant challenge. Traditional scenario generation methods, both data-based and knowledge-based, commonly rely on image data, trajectories, or regulations. However, natural language traffic accident descriptions represent an underutilized source for generating ADS test scenarios. In this work, we introduce a novel method, ScenLaBe, which leverages large language models (LLMs) and behaviour trees to generate test scenarios from these natural language descriptions. Using the NHTSA’s Crashworthiness Data System (CDS) dataset, experiments show that our approach effectively generates critical test scenarios with high behaviour coverage.