In the automotive sector, the zero emissions area has been dominated by battery electric vehicles. However, prospective users cite charging times, large batteries, and the deployment of charging stations as a counter-argument. Hydrogen will offer a solution to these areas, in the future. This research focuses on the development of a prototype three-wheeled vehicle that is named Neumann H2. It integrates state-of-the-art energy storage systems, demonstrating the benefits of solar-, battery-, and hydrogen-powered drives. Of crucial importance for the R&D platform is the system’s ability to record its internal states in a time-synchronous format, providing valuable data for researchers and developers. Given that the platform is equipped with the ROS2 Open-Source interface, the data are recorded in a standardized format. Energy management is supported by artificial intelligence of the “Reinforcement Learning” type, which selects the optimal energy source for operation based on different layers of high-fidelity maps. In addition to powertrain control, the vehicle also uses artificial intelligence to detect the environment. The vehicle’s environment-sensing system is essentially designed to detect, distinguish, and select environmental elements through image segmentation using camera images and then to provide feedback to the user via displays.

Camera and LiDAR sensors in safety-critical autonomous systems suffer undetected distributional shifts that silently corrupt downstream perception. Out-of-distribution (OOD) detection is the established sensor-data-integrity primitive, but no single post hoc detector covers every shift type, and existing detectors lack guarantees on their false-positive rate (FPR). This paper asks when calibrated score fusion helps and provides a distribution-free finite-sample FPR certificate. Four post hoc scores—Maximum Softmax Probability (MSP), Energy, Mahalanobis distance and k-nearest-neighbour (KNN) distance—are calibrated to p-values by the empirical cumulative distribution function (ECDF) and combined by Fisher’s method or cross-backbone z-score averaging, then wrapped in a conformal predictor with Hoeffding-based Probably Approximately Correct (PAC) bounds. On the full-split PUG camera benchmark (215,040 images), uniform same-backbone p-value fusion does not beat the best single detector (Mahalanobis); the gain comes from cross-backbone diversity: a z-score average of Mahalanobis distances over ResNet-50 and frozen DINOv2 reaches a mean area-under-the-ROC-curve (AUROC) of $0.9258$ ($+0.0199$), rising to $0.9292$ ($+0.0233$) with added spectral and dropout signals (DeLong $p<{10}^{-9}$). On the nuScenes LiDAR sensor (256,873 frames), uniform fusion yields only a small, calibration-sensitive gain over the best single detector (MSP), so the substantial fusion gain is confined to cross-backbone averaging on the camera. The distribution-free PAC certificate, by contrast, transfers across both sensors with margins below $1.5\%$ ($0.96\%$ camera, $0.25\%$ LiDAR), giving evidence usable in ISO 26262 and EASA CoDANN safety cases.

Camera-distribution monitoring and semantic out-of-distribution detection assign different meanings to weather and illumination changes. We evaluate specified departures from a camera reference distribution and distinguish the image representation from its scoring rule. Frozen CLIP ViT-B/32 gives AUROC 0.504 on a fixed PUG covariate-shift sample with content-prompt maximum cosine, but an ID-fitted Mahalanobis readout of the same encoder reaches 0.809. On BDD100K night, a text-free CLIP KNN readout reaches 0.913 versus 0.534 for maximum cosine. Prompt/scoring controls further demonstrate readout dependence and the effect of named-condition prior knowledge. On Boreas, the baseline ranks winter and evening shifts in opposite directions (AUROC 0.823 and 0.249). One threshold fitted on 500 separate summer calibration frames misses 44.8% of winter and 99.6% of evening frames, with 5.8% false alarms on held-out summer frames. Scene and position controls partly address acquisition confounding. A preprocessing audit overturns the original adversarial-training benefit; corrected checkpoint results and a separate five-seed study replace that conclusion. These results support evaluating complete encoder–score–prompt–calibration pipelines, with explicit sampling and operational limits.