Mingxuan Liu^1,2,*       Honglin He^1,*       Elisa Ricci^2,3       Wayne Wu¹       Bolei Zhou^{1 1}University of California, Los Angeles     ²University of Trento     ³Fondazione Bruno Kessler ICLR 2026

Introducing UrbanVerse — a system that converts real-world urban scenes from city-tour videos into physics-aware, interactive simulation environments, enabling scalable robot learning in urban spaces with real-world generalization. Unmute for the best audio experience.

Using the extracted scene layout from YouTube videos as a blueprint and assets retrieved from UrbanVerse-100K, UrbanVerse generates simulation environments faithfully grounded in the real-world layout.

For the same city-tour video and its layout, UrbanVerse generates multiple diverse digital cousin scenes by instantiating the layout with different retrieved assets.

For the same city-tour video and its layout, UrbanVerse generates multiple diverse digital cousin scenes by instantiating the layout with different retrieved assets.

View Interactive Category Distribution (Full Screen) →

Object Assets Walkthrough

Rich Per-object Annotation

Road PBRs

Sidewalk PBRs

Sky HDRIs

Real-world city-tour videos from across the globe provide the foundation for urban simulation layout grounding.

Given the uncalibrated RGB city-tour videos, we use the UrbanVerse-Gen pipeline to extract the real-world semantic scene layouts.

@inproceedings{liu2026urbanverse,
  title={UrbanVerse: Scaling Urban Simulation by Watching City-Tour Videos},
  author={Mingxuan Liu and Honglin He and Elisa Ricci and Wayne Wu and Bolei Zhou},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
}