UrbanVerse-Assets-Sim-ready

Simulation-ready USD conversions of the UrbanVerse-100K 3D asset database, for NVIDIA Isaac Sim ≥ 5 / Isaac Lab.

Project: UrbanVerse: Scaling Urban Simulation by Watching City-Tour Videos (ICLR 2026) · Code: VAIL-UCLA/UrbanVerse

What "sim-ready" means here

Each asset was converted from its metric, ground-aligned GLB (adjusted_asset_scaled_bottomed.glb) with Isaac Lab's MeshConverter and validated against the source mesh:

property	value
up axis	Z (glTF's Y-up geometry is rotated +90° about X, so the asset stands upright at identity)
units	metres, metersPerUnit = 1
origin	base of the asset on z = 0
physics	RigidBodyAPI (kinematic), CollisionAPI, MeshCollisionAPI = convexDecomposition
materials	UsdPreviewSurface + textures under textures/

Every stage's world bounds were checked against the exact vertex bounds of its GLB after the axis swap (extents equal to 1e-3, base at z = 0).

Layout

One tar per asset, bucketed by the first two characters of its id, lowercased (329 folders, at most ~460 files each):

usd/
└── <uid[:2].lower()>/
    └── std_<uid>.tar
        └── std_<uid>/
            ├── std_<uid>.usd      # reference this
            ├── config.yaml        # MeshConverterCfg used for the conversion
            └── textures/

<uid> is the UrbanVerse-100K asset id, so annotations from the urbanverse-asset toolkit apply directly. Most ids are 32-character hex; 85 are 26-27-character mixed-case ids, hence the lowercasing. Extracting every tar into one directory gives a flat std_<uid>/ folder per asset.

Usage

Fetch and unpack one asset:

import tarfile
from huggingface_hub import hf_hub_download

uid = "0011e400ad1042cfb6a990376240b091"
tar = hf_hub_download("UCLA-VAIL/UrbanVerse-Assets-Sim-ready", f"usd/{uid[:2].lower()}/std_{uid}.tar",
                      repo_type="dataset")
tarfile.open(tar).extractall("assets")          # -> assets/std_<uid>/std_<uid>.usd

Everything (~1.7 TB):

hf download UCLA-VAIL/UrbanVerse-Assets-Sim-ready --repo-type dataset --local-dir assets_tar
find assets_tar/usd -name '*.tar' -print0 | xargs -0 -n1 -P8 tar -C assets -xf

Then, in Isaac Lab:

import isaaclab.sim as sim_utils

cfg = sim_utils.UsdFileCfg(usd_path="assets/std_<uid>/std_<uid>.usd")
cfg.func("/World/asset", cfg, translation=(0.0, 0.0, 0.0))

Coverage

102,438 of the 102,444 assets in the UrbanVerse-100K GLB corpus (1.78 TB of tars). The remaining 6 (10fe09a8…, 1d4b27a6…, 78ac73b2…, 95f11f5a…, d92aedac…, eceb3109…) are 0-byte GLB files upstream — the public UrbanVerse-100K release only carries their renders — so they could not be converted. Per-batch records, the list of those ids and the ~2,300 assets whose source GLB is not ground-aligned (base above or below z = 0, converted faithfully) are in scenes/assets_batches_state.json. Conversion tooling: scripts/profile_glb_to_usd.py, scripts/convert_assets_batched.py and urbanverse_100k/urbanverse_asset/_glb_to_usd.py in the code repository.

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},
}