UrbanVerse: Scaling Urban Simulation by Watching City-Tour Videos

ICLR, 2026

---

Mingxuan Liu^1,2,*, Honglin He^1,*, Elisa Ricci^2,3, Wayne Wu¹, Bolei Zhou¹ (*Equal contribution)

_{¹University of California, Los Angeles    ²University of Trento    ³Fondazione Bruno Kessler}

---

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.

Click the image below to watch the introductory video.

Updates

• Mar 19, 2026 - CraftBench scenes are released here
• Mar 06, 2026 - UrbanVerse-100K 3D asset database is released here
• Oct 20, 2025 - UrbanVerse paper is now available here

Open-Source Roadmap (priority order)

• UrbanVerse-100K Dataset: release annotations and rescaled assets
• UrbanVerse-100K Dataset Toolkit: release Python package urbanverse-asset
• CraftBench scenes: release the 10 professionally designed test-only scenes
• Scene Toolkit: release Python package urbanverse-scene
• UrbanVerse Scenes Repository: release the 160 simulation-ready urban scenes generated by UrbanVerse
• UrbanVerse-Gen Pipeline: release the real-to-sim automatic scene construction code
• RL Training Pipeline: release in-simulation policy training and evaluation code

Status will be updated as each component is released. If you use UrbanVerse, please cite the paper.

Citation

@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},
}