Physics-Guided Distributional Shaping for Runtime Risk Assessment in Autonomous Driving

Abstract

Reliable risk assessment in interactive traffic remains a fundamental challenge, as rule-based methods rely on rigid safety constraints while pure statistical models lack explicit kinematic boundaries. We propose a physics-guided distributional shaping framework that reweights a learned interaction density with a bounded kinematic energy through Gibbs tilting, and approximates the shaped density with a closed-form Gaussian mixture model (GMM) for efficient online scoring. Unlike runtime rule correction, the proposed framework embeds the safety bias directly into the scoring distribution offline, while a distance-aware activation term further reduces false alarms from distant non-threatening agents. In zero-shot transfer from nuPlan to Lyft under a Time-to-Collision (TTC)-based high-risk proxy, the proposed method achieves 0.9649 receiver operating characteristic area under the curve (ROC-AUC) and 77.98% true positive rate (TPR) at 5% false positive rate (FPR), outperforming rule-based and purely statistical baselines in pairwise risk ranking.