CRASH: Context-Aware Recognition of Agents for Simulation of High‑risk Driving
Minhee Cho, Hayeon Jo, Dongbo Min
AI summary
Problem
Prior safety-critical scenario generation methods rely on infeasible future trajectory predictions and heuristic spatial proximity, producing implausible interactions that misalign with real-world driving dynamics and contextual risk.
Approach
CRASH filters out static agents via Motion-Aware Masking and uses a cross-attention module to probabilistically identify threatening agents based solely on historical and current interactions, jointly optimizing agent selection and trajectory generation.
Key results
- First learnable framework for adversarial agent selection using only past and present observations
- Motion-Aware Masking effectively filters static, non-threatening agents to reduce training noise
- Adversarial agent Selection Module probabilistically estimates collision likelihood via cross-attention
- Achieves ~65% collision success rate on nuScenes and Waymo, significantly outperforming STRIVE and DiffScene
Why it matters
Enables autonomous vehicle developers to efficiently stress-test safety-critical edge cases with realistic, context-aware simulations that align with real-world temporal constraints.
Abstract
Evaluating the safety of autonomous vehicles requires simulation of safety-critical scenarios such as po- tential collisions, which are difficult to reproduce in real- world environments. Prior methods rely on future trajectory predictions and heuristically select adversarial agents based on spatial proximity to the ego vehicle, often producing unrealistic scenarios that misalign with real-world temporal dynamics and contextual risk. To address these issues, we propose CRASH, the first learning-based adversarial agent selection approach that operates solely on past and present observations. It comprises two key components: (1) a Motion-Aware Masking (MAM) module that filters out static agents unlikely to collide with the ego vehicle due to negligible movement, and (2) an Adver- sarial agent Selection Module (ASM) that models contextual interactions to probabilistically estimate each agent’s likelihood of inducing a collision with the ego vehicle. Experiments on the nuScenes and Waymo datasets demonstrate that CRASH significantly improves the success rate of generating realistic collision scenarios under both replay and rule-based planners, validating the effectiveness of context-aware agent modeling without access to future information.