Integrating Data-Driven and Knowledge-Driven Methodologies for Safety-Critical Scenario Generation in Autonomous Vehicle Validation

Autonomous driving systems (ADS) enhance driving safety and efficiency but face significant challenges in safety-critical scenarios, necessitating thorough pre-deployment validation. Traditional knowledge-driven scenario generation methods are detailed yet slow and often insufficient for comprehensive ADS domain coverage. Conversely, data-driven approaches, employing real-world data and deep learning, offer efficiency but consume extensive computational resources and focus mainly on isolated domains. This paper introduces BridgeGen, a novel framework that integrates these methodologies to overcome their limitations. BridgeGen uses an ontology-based approach to convert abstract human knowledge into structured, symbol-based spatial constraints for efficient cross-domain exploration. Extensively tested with the Carla simulator, BridgeGen efficiently generates a diverse array of critical safety scenarios essential for robust ADS validation. Within this framework, we developed a scenario generation toolkit based on the ontology-based concept, assessing single-objective (PPO and PSO) and multi-objective (NSGA-II -DT) optimization algorithms. Our results indicate that PPO and PSO enhance critical scenario generation by 1.61 and 1.41 times, respectively, compared to baseline algorithms. NSGA-II -DT, used for multi-objective optimization, improves search quality and efficiency by 1.22 times, effectively avoiding local optima and ensuring broader scenario coverage.