STRETCH: Generating Challenging Scenarios for Testing Collision Avoidance Systems

Collision avoidance systems are fundamental for autonomous driving and need to be tested thoroughly to check whether they safely handle critical scenarios. Testing collision avoidance systems is generally done by means of scenario-based testing using simulators and comes with the main challenge of generating situations that are realistic but avoidable. In other words, driving scenarios must stress the collision avoidance functionalities while being representative. Existing crash databases and accident reports describe observed accidents and enable to (re)create realistic collisions in simulations; however, as those data sources focus on the impact, their data do not generally lead to avoidable collision scenarios. To address this issue, we propose STRETCH, which generates realistic, critical, and avoidable collision scenarios by extending focused collision descriptions using a multi-objective optimization algorithm. Thanks to STRETCH, developers and testers can automatically generate challenging test cases based on realistic crash scenarios.