Interpretability Boosts End-to-End Autonomous Driving Safety.

The increasing adoption of end-to-end learning in autonomous driving systems introduces significant model complexity and opacity, raising concerns about unintended or incorrect behaviors. This research addresses these issues by integrating an unsupervised dictionary learning module into state-of-the-art driving models. This interpretability module works post hoc to break down complex driving behaviors into semantically meaningful concepts. It then demonstrates the causal influence of these concepts on the model's driving decisions, effectively revealing the underlying logic behind trajectory predictions. Furthermore, the framework allows for targeted interventions at the concept level. By manipulating these concepts, researchers can correct erroneous driving decisions, leading to measurable improvements in overall driving performance. This approach highlights how interpretability can reduce model opacity, identify flaws, and enable precise mitigation, ultimately enhancing the safety and reliability of autonomous vehicles.