Can Attribution Predict Risk? From Multi-View Attribution to Planning Risk Signals in End-to-End Autonomous Driving
Researchers have developed a new hierarchical attribution framework designed to predict risks in end-to-end autonomous driving models. This method analyzes visual inputs across multiple camera views to identify critical regions and their influence on trajectory generation. The framework extracts three key statistics—attribution entropy, spatial variance within cameras, and a cross-camera Gini coefficient—which correlate with planning risks like trajectory errors and potential collisions. AI
IMPACT Introduces a novel method for risk prediction in autonomous driving systems, potentially improving safety and reliability.