Post by Journal of Systems and Software

Simulation testing can produce unrealistic adversarial NPC behaviors. Tiexin Wang, Shuo Tian, Gulent Asalif Minas, and Chunyang Bian propose AMACollision, a multi-agent framework for generating realistic, rule-compliant adversarial driving scenarios. The framework applies multi-agent reinforcement learning with policies that integrate multi-modal sensor fusion and temporal decision-making, and introduces a two-stage reward mechanism designed to balance adversarial objectives with adherence to traffic regulations. In experiments that integrated AMACollision with a high-fidelity simulator to test two ADSs across three road layouts, the approach indicated improved scenario realism and higher generation efficiency compared with a state-of-the-art method, based on seven evaluation metrics (for example, collision rate). Read the full paper at: https://lnkd.in/dX5QWm9Q #AutonomousDriving #SimulationTesting #ReinforcementLearning #SafetyEngineering