CADET: Physics-Grounded Causal Auditing and Training-Free Deconfounding of End-to-End Driving Planners

CADET is a novel framework designed for the causal auditing and deconfounding of end-to-end (E2E) driving planners, addressing their vulnerability to statistical shortcuts that can lead to unreliable decision-making in autonomous driving scenarios. Unlike existing methods that necessitate retraining large models, CADET operates without parameter updates, allowing for the auditing and correction of spurious reliance in pretrained planners. This capability is crucial for practitioners as it enhances the reliability of E2E systems in complex driving environments without the overhead of model retraining.