UrbanVerse-100K Dataset

UrbanVerse-100K is a large-scale, physics-aware 3D asset and material database curated for urban simulation, physical and embodied AI research. It contains over 102K metric-scale urban object assets (GLB/USD), along with 646 4K sky maps (HDR) and 403 4K ground (road/sidewalk/terrain) materials (MDL), each annotated with rich semantic and physical attributes. The dataset is IsaacSim-ready, enabling scalable construction of realistic urban simulation environments for training and evaluating embodied AI agents, and beyond.

Current release version: 0.1.0

Continuous improvement discussions: If you have questions, suggestions for new features, or ideas for adding new assets, please post them in the corresponding discussion threads below:

• 💡 urbanverse-asset API related thread
• 🧱 Assets related thread

Total size: 1.18 TB

Assets overview:

Asset Type	Count	Format	Description
3D Objects	102,445	.glb	Metric-scale urban 3D objects in GLB format across 659 categories
Object Annotations	102,445	.json	Per-asset annotation
Sky Maps	646	.hdr	HDR environment maps
Road Materials	98	.mdl	PBR road surface materials
Sidewalk Materials	190	.mdl	PBR sidewalk surface materials
Terrain Materials	115	.mdl	PBR terrain surface materials

---

Get Started

UrbanVerse-100K Dataset Usage Toolkit: We provide a Python toolkit urbanverse-asset for easy, robust downloading and exploring the full dataset, including searching/retrieving, downloading, format conversion, and interactive 3D visualization.

Request Access: First, you need to get access to this repository approved. Visit the repository page, fill out the form and click "Submit". Once your request is approved, you will be able to use our toolkit or any other method you prefer to download and access the files.

Install & Authentication

Hardware Requirements: Windows, macOS, and Ubuntu are all supported for using most APIs (e.g., download, search, and viewer), as the built-in viewer is based on Three.js.

pip install urbanverse-asset

Authenticate your machine with your Hugging Face account that have access to this HF repository approved:

hf auth login

You might be prompted to paste your Hugging Face access token, if you are not logged-in before. You can create a token here: https://huggingface.co/settings/tokens

After login, verify the current account:

hf auth whoami

Optional — for GLB-to-USD conversion API (requires NVIDIA IsaacSim and IsaacLab):

Hardware Requirements for IsaacSim: A Windows or Ubuntu machine equipped with an NVIDIA GeForce RTX GPU is needed for using IsaacSim based glb-to-usd conversion function to make assets IsaacSim-ready.

pip install 'isaacsim[all,extscache]==4.5.0' --extra-index-url https://pypi.nvidia.com
git clone https://github.com/OatmealLiu/IsaacLab.git
cd IsaacLab && ./isaaclab.sh --install

Explore

We recommend using the following documentation to quickly and efficiently get started with downloading and using UrbanVerse-100K.

• Interactive walkthrough: demo notebook
• API reference: urbanverse-asset README

Full Download (see the next section for selective downloading)

We recommend downloading the dataset using our API rather than cloning the repository, as it provides more robust downloading and reliable resuming support.

import urbanverse_asset as uva

# (Optional) set a custom cache directory (default: ~/.cache/urbanverse/)
uva.set("~/datasets/urbanverse")

# WARNING: This downloads the full dataset (~1.18 TB).
#          See selective and flexible download under user-defined conditions using our API in the following section.
uva.download_all(num_workers=32)

This will download the entire dataset (~1.18 TB) to the specified cache directory, which may take a considerable amount of time. As a reference, during our stress test with a 500 MB/s internet connection, the full download took approximately 6–8 hours.

If the download is interrupted for any reason, you can simply rerun the function above to resume the process, or use the repair function below. Both methods allow the download to reliably continue from the point of interruption. They are designed to support robust resuming.

# Output a report of the integrity of the downloaded files, including the number of complete and missing files.
report = uva.check_integrity()
print(f"Complete: {report['complete']}, Missing: {report['total_missing']}")

# Automatically download only the missing or incomplete files
uva.repair(num_workers=32)

Asset Filtering and Selective Download (⭐highly recommended⭐)

See the detailed tutorial notebook here and the API guide here.

You can also download specific subsets instead of the full dataset via filtering:

# Search for specific assets by text query
uids = uva.object.get_uids_conditioned(
    categories=["vehicle"],
    query="Old yellow Italian style two-door car",
    top_k=5,
)

# Download only the selected assets
result = uva.object.load(uids)

# Browse in an interactive 3D viewer
uva.viewer.object_show(uids)

Material assets can be searched and downloaded in the same way:

# Sky materials
sunny = uva.sky.get_descriptions_conditioned(query="Sunny summer day", top_k=10)
uva.sky.load_materials(sunny)
uva.viewer.sky_show(sunny)

# Road materials
highway = uva.road.get_descriptions_conditioned(query="Highway", top_k=10)
uva.road.load_materials(highway)
uva.viewer.road_show(highway)

# Sidewalk materials
stone = uva.sidewalk.get_descriptions_conditioned(query="Bumpy stone road", top_k=10)
uva.sidewalk.load_materials(stone)
uva.viewer.sidewalk_show(stone)

# Terrain materials
grass = uva.terrain.get_descriptions_conditioned(query="Grassland", top_k=10)
uva.terrain.load_materials(grass)
uva.viewer.terrain_show(grass)

---

License Distribution

While the overall dataset collection is provided under the ODC-BY license, the individual assets within the dataset fall under the following specific licenses, as attributed in each per-asset annotation file.

License	Asset Count	Permissions & Restrictions
CC BY	83,858	Safe for redistribution with attribution (commercial use allowed)
CC BY-NC-SA	10,437	Non-commercial use only + must share alike
CC BY-NC	5,824	Non-commercial use only
CC BY-SA	1,365	Must give credit + share alike
CC0	960	Public domain (no restrictions)

---

Citation

If you use this dataset or find our work helpful, please cite our paper:

@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},
}

---

Credits and Acknowledgements

Our 3D dataset is curated from Objaverse, which aggregates 3D assets from Sketchfab, and is further annotated and organized by our UrbanVerse team. Further, we source free-license HDR sky maps from Poly Haven, and obtain PBR road, sidewalk, and terrain materials from AmbientCG. We sincerely thank the authors, designers, and artists who openly share these high-quality assets, which make this research possible.