Adaptive risk inversion with iterative exploration for high-risk AV-VRU interactions.

Area of Science:

• Automotive Safety Engineering
• Artificial Intelligence in Transportation
• Human-Robot Interaction

Background:

• Automated vehicle (AV) safety validation requires realistic high-risk scenarios involving vulnerable road users (VRUs).
• Current methods struggle to capture VRU behavioral complexity and explore rare, critical long-tail risks.
• This gap limits the rigor and representativeness of AV safety evaluations.

Purpose of the Study:

• To introduce ARISE (Adaptive Risk Inversion for Scenario Exploration), a novel framework for generating long-tail, high-risk AV-VRU interaction scenarios.
• To systematically identify, parameterize, and refine high-risk regions in the scenario space.
• To improve the safety validation of AVs in complex interactions.

Main Methods:

• ARISE integrates failure-driven feedback with interpretable behavioral modeling.
• A multidimensional VRU behavioral model captures decision-making variability.
• The framework was implemented on the OnSite simulation environment for iterative scenario generation.

Main Results:

• ARISE significantly enhances the diversity and severity of generated high-risk AV-VRU interactions compared to conventional methods.
• The study identified critical risk intervals with clear interpretability.
• The framework facilitates systematic evaluation of AV safety performance.

Conclusions:

• ARISE provides a robust methodological foundation for AV safety validation in challenging VRU interactions.
• The framework effectively bridges the gap between scenario generation and safety assessment.
• This approach is crucial for ensuring the safety of AVs operating alongside VRUs.