AMACollision: An advanced framework for testing autonomous vehicles based on adversarial multi-agent

Autonomous driving, as one of the typical safety-critical domains, requires strict safety evaluation. Simulation testing, due to the advantages such as high efficiency and low cost, is an important enabling means of evaluation. Currently, researches toward simulation testing of Autonomous Driving Systems (ADSs) mainly aim at identifying safety-critical driving scenarios. However, the realism of the generated scenarios and the generation efficiency remain as two challenges. Therefore, we propose AMACollision, a multi-agent-based framework for generating highly realistic critical driving scenarios. AMACollision employs a deep neural network that integrates multi-modal sensor fusion and temporal decision-making, enabling robust scene understanding and efficient training of the multi-agent that act as Non-Player Characters (NPCs). To enhance the realism of the generated scenarios, a two-stage reward mechanism, which ensures NPCs’ behaviors comply with traffic regulations, is introduced. To evaluate the performance of AMACollision, we integrate it with a high-fidelity simulator and conduct extensive experiments testing two ADSs on three various road structures. Experimental results, collected from seven metrics, e.g., collision rate, demonstrate that AMACollision outperformed the state-of-the-art method in both the realism of the generated scenarios and the generation efficiency.