Safety Concerns in Autonomous Driving: A Taxonomy of Test and Evaluation Approaches

Ensuring the safety of autonomous vehicles (AVs) requires rigorous and scalable testing methodologies capable of capturing both routine and safety-critical scenarios. Scenario-based testing has emerged as a vital approach to expose AVs to diverse and challenging conditions beyond traditional road mileage accumulation. This survey focuses on scenario generation—an essential component enabling automated, efficient, and comprehensive testing of autonomous driving systems (ADS). We categorize existing scenario generation methods into three primary paradigms: rule-based, data-driven, and learning-based. For each, we analyze the core methodologies, simulation platforms, scenario description languages, and evaluation metrics used to assess realism, diversity, and criticality. We further identify key research challenges such as the reality gap, limited data generalization, and rare-event modeling, and discuss emerging trends including language-driven generation, hybrid modeling frameworks, and standardized scenario repositories. This work provides a unified perspective on scenario generation, aiming to support researchers and practitioners in advancing safe and certifiable autonomous driving technologies.