Model-Based Critical Pre-Crash Interaction Identification via Multi-Agent Collaboration

Historical pre-crash scenarios contain a sequence of vehicle-to-vehicle interactions, among which the critical pre-crash interaction is the one that most directly leads to the eventual collision. Identifying such critical interactions is important for uncovering potential safety risks in Autonomous Driving Systems (ADSs). However, this task remains challenging because the critical interaction is often embedded within numerous preceding interactions and needs to be distinguished by considering how the pre-crash scenario evolves over time. To address this challenge, we propose MICA, a model-based multi-agent approach for identifying critical pre-crash interactions from textual accident reports. Unlike existing LLM-based approaches that directly identify critical interactions from accident reports, MICA explicitly models the evolution of pre-crash interactions through multi-agent collaboration. Specialized LLM-based agents extract and validate relevant information to construct a Pre-crash Evolution Model, from which the critical interaction is identified by matching the modeled interactions to predefined pre-crash patterns. We evaluate MICA on 200 accident reports released by the National Highway Traffic Safety Administration (NHTSA) and the California Department of Motor Vehicles (DMV CA). The results demonstrate that MICA effectively identifies critical pre-crash interactions across different accident report data sources, scenario complexities, and underlying LLMs.