Safety-Critical Scenario Generation for Self-Driving Systems Based on Domain Models

The end-to-end automated driving system (ADS) has the capability to eliminate human errors present in driving tasks, thereby reducing the occurrence of traffic accidents and enhancing safety in the transportation sector. However, due to the complexity and opacity of its internal decision-making processes, which are often perceived as a black box, there is a potential for incorrect or unexpected extreme situations to arise, necessitating stringent robustness requirements. To comprehensively evaluate the performance of ADS, it is essential to generate a diverse array of safety-critical scenarios. This paper proposes a method aimed at generating test scenarios for ADS safety by combining domain models with hazardous factors that lead to violations or accidents. The method derives key objects and their environmental conditions that may cause accidents or violations from the hazardous factors obtained from corner cases. Based on this, we perform traversal operations on the domain model to create variations or reproductions of the identified objects and environmental conditions. Ultimately, through this series of operations, the generated safety-critical scenarios not only closely match real-world situations but also provide a reliable basis for further safety assessments. We simulated the generated scenarios on the industrial-grade ADS platform Apollo, successfully identifying safety violations.