WildCity: A Real-World City-Scale Testbed for Rendering, Simulation, and Spatial Intelligence

WildCity introduces a real-world multimodal, city-scale dataset collected by autonomous fleets in complex urban environments. It contains 18 trajectories averaging 83.7 kilometers each, while retaining in-the-wild difficulties such as dynamic objects, changing illumination, and imperfect camera poses. The authors also provide an urban-oriented reconstruction baseline and turn reconstructed environments into a closed-loop simulator. The paper analyzes scalability, extrapolation, and uncertainty as central obstacles to simulation-ready urban digital twins.