High-risk scenario generation for AV–VRU interaction: An adaptive risk-inversion and scenario-exploration framework

Yujia Zhao^1,2, Ying Ni^1,2, Jialin Fan^1,2, Jie Wang^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 interaction between vulnerable road users (VRUs) and automated vehicles (AVs) represents a critical aspect of AV safety evaluation, especially under realistic high-risk scenarios. However, existing scenario-based testing methodologies often inadequately capture VRUs’ behavioral complexity and struggle to systematically explore rare yet critical long-tail risks. This limitation diminishes the representativeness and rigor of safety validation, potentially underestimating the risks inherent in complex AV–VRU interactions. To address these shortcomings, this study proposes ARISE (Adaptive Risk Inversion for Scenario Exploration) specifically designed for generating long-tail, high-risk AV–VRU interaction scenarios. The framework integrates failure-driven feedback with interpretable behavioral modeling, systematically identifying, parameterizing, and refining high-risk regions within the scenario space. A multidimensional behavioral model capturing VRU decision-making variability is developed, enabling hierarchical extraction of critical single- and multi-parameter risk intervals with clear interpretability. The proposed ARISE is implemented on Open Natural Driving Intelligence Automotive Simulation Test Environment (OnSite), facilitating iterative scenario generation and systematic evaluation of AV safety performance. Experimental case studies validate that ARISE significantly improves both the diversity and severity of generated high-risk interactions compared to conventional methods. These findings underscore ARISE’s capability to effectively bridge the gap between scenario generation and AV safety validation, providing a robust methodological foundation for assessing AVs in challenging and realistic VRU interactions.