UCLA-VAIL/UrbanVerse-CraftBench-Sim-Ready

UrbanVerse-CraftBench Scenes (Sim-Ready)

The 12 hand-crafted UrbanVerse-CraftBench test scenes, upgraded to load correctly in NVIDIA Isaac Sim ≥ 5. Same scenes, same packaging (Collected_export_version.tar per scene, open Collected_export_version/export_version.usd), same license as the original.

What changed, and why

The scenes were exported from Isaac Sim 4.5 with road / sidewalk vMaterials whose inputs:texture_scale was authored as an int while the MDL declares float2.

• Isaac Sim 4.5 rejected that value ([UsdToMdl] Tried to assign a 'int'(USD) to a 'float2'(MDL)) and silently fell back to the material's default tile size - that is what the preview_*.png images show.
• Isaac Sim 5.x rejects it too, but the surface then renders flat black/white.

Every such attribute was re-authored as float2 with the MDL parameter's own default (Mortar 1.0, Paving_Stones 0.5, Rough_Gravel 2.0, ...), reproducing the 4.5 appearance in 5.x. Attributes that were already a valid float2 were left untouched; geometry, layout, assets, textures and physics are unchanged.

Tooling and per-scene provenance (every changed layer with its sha256): scripts/upgrade_isaacsim5.md, scenes/craftbench_simready_manifest.json.

Layout

scene_<id>_<place_type>_<topology>_<attributes>/
├── Collected_export_version.tar   # extract, then open Collected_export_version/export_version.usd
├── cam0_to_world.txt              # per-frame camera-to-world matrices of the preview flythrough
├── preview_front.png / preview_topdown.png / preview_closeup.png
└── preview_video.mp4

Usage

import urbanverse_scene as uvs          # pip install urbanverse-scene
uvs.craftbench.load_scenes(uvs.craftbench.get_descriptions()[:1])

or extract a tar and open export_version.usd in Isaac Sim 5.x.

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},
}