The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset

Researchers at Karlsruhe Institute of Technology and FZI Research Center for Information Technology introduce KITScenes Multimodal, a European autonomous-driving dataset designed to test spatial reasoning in difficult urban environments. Its synchronized robotaxi platform combines 72.5 Mpx global-shutter cameras, seven lidars reaching a measured maximum range of 409.2 m, three 4D imaging radars, and redundant GNSS/INS localization. Covering 62 km2 across Karlsruhe, Frankfurt, and Sindelfingen, the release pairs sensor data with reprojection-accurate Lanelet2 HD maps that encode road topology, regulatory elements, and 3D traffic lights, signs, and poles, and are validated in the open-source Autoware stack. Four benchmarks target complete online HD-map construction, long-range monocular depth, map-grounded novel-view synthesis, and multimodal end-to-end driving. The results expose failures hidden by conventional datasets: beyond 200 m, UniDAC achieves the strongest tested depth estimate but only a δ1 score of 1.78, while current novel-view synthesis loses more than 80 percent of traffic-sign recall at lateral offsets of ±3 m. In open-loop driving, the strongest Epona configuration reaches a 98.3 percent collision-free rate at the 3-second horizon. KITScenes is smaller than large corpora such as NVIDIA’s PhysicalAI AV, but prioritizes sensor fidelity, geographic diversity, complete maps, and deployment-oriented evaluation for Level 4 autonomy.