Abstract
Autonomous driving has the potential to significantly enhance productivityand provide numerous societal benefits. Ensuring robustness in thesesafety-critical systems is essential, particularly when vehicles must navigateadverse weather conditions and sensor corruptions that may not have beenencountered during training. Current methods often overlook uncertaintiesarising from adversarial conditions or distributional shifts, limiting theirreal-world applicability. We propose an efficient adaptation of an uncertaintyestimation technique for 3D occupancy prediction. Our method dynamicallycalibrates model confidence using epistemic uncertainty estimates. Ourevaluation under various camera corruption scenarios, such as fog or missingcameras, demonstrates that our approach effectively quantifies epistemicuncertainty by assigning higher uncertainty values to unseen data. We introduceregion-specific corruptions to simulate defects affecting only a single cameraand validate our findings through both scene-level and region-levelassessments. Our results show superior performance in Out-of-Distribution (OoD)detection and confidence calibration compared to common baselines such as DeepEnsembles and MC-Dropout. Our approach consistently demonstrates reliableuncertainty measures, indicating its potential for enhancing the robustness ofautonomous driving systems in real-world scenarios. Code and dataset areavailable at https://github.com/ika-rwth-aachen/OCCUQ .