Reliable testing of Autonomous Driving Systems (ADS) is essential before deployment on public roads. Real-world testing is considered the most trustworthy validation method because it operates in the ADS's target domain. However, it is expensive, safety-critical, and difficult to reproduce. 2D driving datasets provide realistic traffic scenarios, but they support only model-level evaluation and cannot capture system-level failures that arise from interactions between the ADS and the environment. In contrast, simulators enable controlled and reproducible system-level testing, but they suffer from the reality gap: scenarios are often simplified, and game-engine visuals differ from the real-world data used to train the ADS.