The KI Delta Learning project focused on advancing the robustness and safety of autonomous driving perception systems. Central to the effort was the development of a new feature‑space comparison technique that incorporates structural information such as variance and cross‑correlation rather than relying solely on pixel‑wise distances. Experiments on the KITTI and nuScenes LiDAR datasets showed that this approach produced feature representations that were qualitatively closer to the teacher network and yielded a significant improvement in detection accuracy. Quantitatively, the proposed method achieved a mean error of 0.029 on KITTI, compared with 0.044 for the baseline, and reduced the overall error from 2.864 to 2.825. On the NYU depth dataset the method lowered the error from 0.066 to 0.061 and improved the depth estimation metric from 0.267 to 0.252. These results demonstrate that incorporating structural statistics into the loss function can enhance both classification and depth‑prediction tasks.

Another key contribution was the KING framework, a gradient‑based optimisation technique for generating safety‑critical traffic scenarios. Starting from a benign simulation, KING iteratively perturbs the actions of surrounding vehicles to induce a collision with the ego‑agent. Fine‑tuning the ego‑agent’s neural controller on these adversarial scenarios reduced the collision rate by a noticeable margin, as illustrated in the project figures. This method provides a systematic way to stress‑test perception and planning modules under rare but critical conditions.

The project also addressed domain adaptation between heterogeneous LiDAR setups. A CycleGAN‑based sensor‑to‑sensor translation was trained to map point clouds from the KITTI configuration to the nuScenes configuration, thereby enabling models trained on one dataset to generalise to the other. Visualisations of the translated scans showed that the network preserved semantic structure while adapting to the differing point densities.

For semantic segmentation, the team applied a hierarchical annotation scheme to the SemanticKITTI dataset, extending the original leaf‑node labels with meta‑ and binary classes. The resulting confusion matrix revealed that the hierarchical model achieved higher precision on fine‑grained classes while maintaining overall accuracy. The segmentation network also produced pixel‑wise confidence maps, which were used to estimate uncertainty in the presence of occlusions or shadows.

Two peer‑reviewed publications emerged from the project. The first paper introduced a method for detecting inputs that lead to unreliable predictions, enabling selective intervention in autonomous driving pipelines. The second paper presented a pixel‑wise uncertainty estimation technique that accounts for image noise, providing a more reliable confidence measure for downstream decision making.

Collaboration was organised around a consortium led by Mercedes‑Benz AG, which supplied real‑world automotive data, evaluation benchmarks, and domain expertise. Partner research institutions developed the algorithms, performed simulations in the CARLA environment, and conducted the empirical evaluations. The project followed a structured plan that was updated during the course of the work, with a total duration of approximately 24 months. Funding was provided through a German research programme aimed at advancing AI safety in automotive applications. The final partner‑specific report summarises the technical achievements, outlines the economic and scientific impact, and highlights the potential for integration into commercial autonomous driving stacks.