Study on Data-Driven Scenario Construction for Autonomous Driving Testing

Guaranteeing autonomous vehicle (AV) safety requires scalable evaluations that capture both everyday driving and critical edge cases. Scenario-based testing has become a key strategy, with scenario generation serving as its core. We classify generation techniques into three categories: rule-oriented, data-driven, and learning-enabled. For each, we examine representative approaches, simulation tools, description standards, and assessment metrics addressing fidelity, diversity, and risk exposure. Persistent issues include the simulation–reality gap, limited transferability, and the challenge of modeling rare events. Emerging directions such as language-guided generation, hybrid architectures, and open scenario repositories point toward more reliable and certifiable testing of AV systems.