Making autonomous vehicle perception systems explainable and trustworthy

Deep neural networks power autonomous driving perception but their opacity conflicts with safety requirements and regulatory expectations. This work bridges that gap by building a transformer-based 3D object detector that provides faithful explanations via attention mechanisms, calibrated uncertainty estimates, and improved robustness. The team validated explanation faithfulness through perturbation tests, integrated uncertainty calibration, and applied robustness-enhancing training. The system was deployed in a prototype vehicle with a live interface showing saliency maps and model confidence, demonstrating practical feasibility for trustworthy perception monitoring during operation.