CriSGen: Constraint-based Generation of Critical Scenarios for Autonomous Vehicles

Andreas Nonnengart, Matthias Klusch, Christian Müller

In: Proceedings of International Workshop on Formal Methods for Autonomous Systems. International Workshop on Formal Methods for Autonomous Systems (FMAS-2019) located at 3rd World Congress on Formal Methods October 7-11 Porto Portugal Springer 2019.


Ensuring pedestrian-safety is paramount to the acceptance and success of autonomous cars. The scenario-based training and testing of such self-driving vehicles in virtual driving simulation environments has increasingly gained attention in the past years. A key challenge is the automated generation of critical traffic scenarios which usually are rare in real-world traffic, while computing and testing all possible scenarios is infeasible in practice. In this paper, we present a formal method-based approach CriSGen for an automated and complete generation of critical traffic scenarios for virtual training of self-driving cars. These scenarios are determined as close variants of given but uncritical and formally abstracted scenarios via reasoning on their non-linear arithmetic constraint formulas, such that the original maneuver of the self-driving car in them will not be pedestrian-safe anymore, enforcing it to further adapt the behavior during training. (pdf, 524 KB )

German Research Center for Artificial Intelligence
Deutsches Forschungszentrum für Künstliche Intelligenz