Breaking survivorship bias in driving world models

This research addresses a critical flaw in autonomous driving world models where they overestimate safety due to biased training data that excludes accidents. The authors propose a Neutral World Model (NWM) trained on a mixture of routine driving and real accident videos using a bounding-box-guided objective. This approach allows the model to faithfully anticipate imminent collisions while maintaining low false collision rates and competitive quality in routine scenarios. Evaluated on challenging NAVSIM benchmarks, the NWM selects better trajectories more reliably than baseline models. Zero-shot experiments with real-world pedestrian dummy crossing data collected by an autonomous research vehicle further demonstrate successful transfer to unseen sensor observations, confirming that correcting survivorship bias improves safety-critical decision-making.