Search-Based Testing and Debugging for Autonomous Driving Systems in Simulated Environments

To rigorously test the behaviour of an autonomous driving system (ADS), it is essential to consider a driving scenario, i.e., a sequence of scenes, each of which consists of various scenario entities, such as surrounding vehicles, pedestrians, road layouts, traffic lights, and weather conditions, that might affect the driving behaviour.

Considering the high cost and risk involved in testing ADSs in the real world, high-fidelity driving simulations have been increasingly used to verify the safety and reliability of ADS.

For example, many approaches have been proposed in the last few years to automatically generate test scenarios that can reveal the unknown failures of a given ADS (e.g., colliding with other vehicles and pedestrians).

However, they largely suffer from scalability due to the vast number of possible scenario entities that would be included in a test scenario.

Furthermore, existing approaches remain at the testing stage (i.e., finding failure-inducing scenarios) without proper debugging (i.e., localising the root cause of the failures).