Safety-Violation Scenarios Search for ADS via Multi-Objective Genetic Algorithm

Simulation testing of Autonomous Driving Systems (ADS) has become prominent, but prevalent techniques often miss bugs due to insufficient cross-layer integration and employ simplistic participant models, leading to a mismatch with real-world complexities and flawed ADS evaluation. The paper presents a Multi-Objective Genetic Algorithm (MOGA)-based method for creating scenarios that lead to safety violations by the ego vehicle. It considers complex conditions with various actor types and behaviors, using Behavior Trees for actor trajectories. The method includes a multi-dimensional metric to optimize MOGA for critical scenarios and reduce irrelevant ones, considering the ego vehicle’s performance and the environment. Tests show the method produced $\mathbf{1, 5 5 2}$ scenarios in 24 hours, identifying 53 that led to $\mathbf{1 0}$ different safety violations, proving the effectiveness of behavior tree actors in increasing complexity. It also outperforms traditional methods in identifying a wider range of violations faster and triples the occurrence of critical safety scenarios.