The Autotech.agil project, carried out by Intel Deutschland GmbH and funded under the German “Elektronik und Softwareentwicklungsmethoden für die Digitalisierung der Automobilität (MANNHEIM)” program (funder code 01IS22088I), ran from 1 October 2022 to 6 February 2023. The report, dated 1 December 2023, documents the shortened effort that was curtailed by personnel changes at Intel. The project’s goal was to develop an open, standardized architecture for future mobility systems, with a particular focus on integrating uncertainty quantification into the vehicle’s decision‑making pipeline for collision avoidance and route planning.

Technically, the work explored several complementary uncertainty‑estimation techniques. Monte‑Carlo dropout was employed to randomly deactivate connections or neurons during inference, producing slightly different outputs on each pass and thereby providing a simple, sample‑free estimate of model uncertainty. Deep ensembles, which run multiple independently trained networks in parallel, were also investigated; the ensemble members were sometimes generated by applying dropout during training, combining the strengths of both approaches. Evidential deep learning was used to model uncertainty explicitly through subjective logic, representing belief assignments as Dirichlet distributions. Learned confidence estimates were introduced by augmenting object‑detection networks with additional output vectors that predict the model’s confidence, reducing loss when predictions are correct and improving calibration. Deterministic uncertainty quantification leveraged radial‑basis‑function networks with a gradient penalty to enforce sensitivity to input changes, enabling reliable detection of out‑of‑distribution samples by measuring distance to class centroids. Finally, aleatoric uncertainty was addressed by exploiting redundancy in modern object detectors: multiple detections of the same object allow the calculation of probability distributions for bounding‑box position and size in a single inference pass.

The project evaluated these methods using a suite of calibration and accuracy metrics. Reliability diagrams were generated to compare predicted probabilities with observed frequencies. Expected calibration error (ECE) quantified the average discrepancy between predicted confidence and actual accuracy, guiding model calibration. The Brier score, a proper scoring rule for binary outcomes, was used to assess the quality of probabilistic predictions. For regression tasks, a retention curve illustrated the relationship between prediction correctness and uncertainty, while negative log‑likelihood served both as a loss function and a performance metric for probabilistic models. These metrics were applied to both classification and regression problems, demonstrating the versatility of the proposed uncertainty‑aware architectures.

In terms of collaboration, Intel Deutschland GmbH led the technical development and integration of the uncertainty modules into a unified software stack. The project was part of a broader effort to standardize interfaces across automotive sensors, including camera and LiDAR modalities, and to embed uncertainty information into higher‑level decision modules. While the report does not list additional university partners explicitly, the project’s alignment with the MANNHEIM program implies coordination with academic research groups focused on automotive electronics and software development. The early termination of Intel’s participation meant that the final deliverables were limited to the conceptual framework and preliminary experimental results presented in the report.

Overall, the Autotech.agil project advanced the state of the art in uncertainty quantification for autonomous driving by combining dropout, deep ensembles, evidential learning, learned confidence, deterministic RBF‑based methods, and redundancy‑based aleatoric estimation. The resulting framework offers a practical pathway to embed calibrated uncertainty into vehicle decision‑making, thereby enhancing safety in collision avoidance and route planning scenarios.