Uncertainty-Aware LiDAR-Camera Autonomy via Conformal Prediction and Principled Abstention

Safety-critical perception systems demand both reliable uncertainty quantification and principled abstention to ensure robustness under diverse conditions. We propose a novel conformalization framework that delivers statistically guaranteed uncertainty estimates and enables selective prediction in high-risk scenarios. Our method integrates a conformal threshold for valid prediction sets with an abstention threshold optimized via ROC analysis, offering distribution-free coverage guarantees (≥ 1−α) while flagging unreliable outputs. Extensive evaluations on CIFAR-100, ImageNet1K, and ModelNet40 demonstrate superior robustness across camera and LiDAR modalities under varying perturbations. The framework maintains high detection performance (AUC: 0.993→0.995), preserves high coverage (>90.0%), and supports adaptive abstention (13.5% → 63.4% ± 0.5) as environmental severity increases. For LiDAR, it sustains strong coverage (>84.5%) while abstaining effectively on uncertain inputs. Under heavy perturbations, our method remains stable (AUC: 0.995 ± 0.001), consistently outperforming baselines across all modalities. This unified approach closes the gap between theoretical guarantees and real-world deployment needs.