Interactive Scenario Generation for Autonomous Vehicle Testing: Vehicle-Cyclist Group Interactions

Ying Ni^1,2, Yujia Zhao^1,2, Jianqiang Li^1,2, Jialin Fan^1,2, and Jian Sun^1,2

¹ College of Transportation, Tongji University, Shanghai, China

² Key Laboratory of Road and Traffic Engineering, Ministry of Education, Shanghai, China

The rapid development of autonomous vehicles (AVs) has highlighted the urgent need for virtual testing to validate their safety in complex real-world traffic scenarios. Interactions at intersections involving groups of cyclists represent a particularly challenging scenario for AVs, requiring the generation of realistic and diverse vehicle-cyclist group interaction scenarios to assess AV performance. However, most existing research relies on rule-based models for scenario generation and overlooks the leader-follower effect within cyclist groups, neglecting the correlation between cyclists’ decisions. This results in generated scenarios that lack interactivity and realism. In this paper, we propose a framework based on association strategy learning to generate vehicle-cyclist group interactive scenarios. This framework captures both the spatiotemporal information of individuals within the group and the decision-making sequences related to the leader-follower dynamics in group interactions, thus enabling the generation of interpretable interaction scenarios. Experimental results demonstrate that the scenarios generated by this model outperform the baseline model in terms of coverage, realism, and risk, fulfilling the comprehensive testing requirements of autonomous driving systems and contributing to a more thorough evaluation of AV performance.