imagine-io/PhysicalAI-Simready-Kitchen-Assets

PhysicalAI SimReady Kitchen Assets

The source asset library behind the PhysicalAI SimReady Kitchens scene datasets: the textures, environment maps and SimReady models that the generator composes into procedurally-varied kitchen scenes.

This repo holds the ingredients, not the scenes. For the generated scenes themselves, see the companion dataset imagine-io/PhysicalAI-SimReady-Kitchens-v1.

Why this repo exists

Generated scenes used to be fully self-contained — every texture a scene touched and every SimReady package it placed was copied into the scene's .usdz. That makes a scene portable but very large (~700 MB each), and it duplicates the same 127 MB HDRI and the same 141 MB floor normal map across thousands of scenes.

Publishing the assets once, here, lets each scene reference them by URL instead:

def Xform "Fryer" (
    prepend references = @https://huggingface.co/datasets/imagine-io/PhysicalAI-Simready-Kitchen-Assets/resolve/main/Models/appliances/Air_fryer_1.usdz[air_fryer_1.usd]@
)
{
}

Layout

The repo mirrors the generator's asset tree verbatim — the path here is the path on disk.

Path	Size	Contents
Materials/	~2.41 GiB	246 PBR texture maps (.jpg / .png) + 37 material.json parameter files
Models/	~2.89 GiB	453 .glb source meshes + 102 .usdz SimReady packages
hdri/	~765 MiB	10 environment maps (.exr / .hdr), 2K–8K

Materials/ has two shapes: 37 named-material folders (material.json plus 2–3 maps) and 4 texture pools — floor/, wall/, ceiling/, wall_tiles_textures/ — whose leaves are <name>_<size>_meter/ folders carrying the physical tiling scale in the directory name.

Models/ is flat within each category: appliances/, backsplash/, cabinets/, ceiling/, clutters/, doors/, faucets/, handles/, props/, slabs/. In appliances/ and clutters/ every .glb has a matching .usdz twin; the .usdz is the SimReady package (physics APIs, collision approximations, MDL materials) and the .glb is the mesh source.

SimReady packages

Each .usdz is a stored (uncompressed) zip holding a root .usd, a gltf/pbr.mdl, and SubUSDs/textures/*. The root member is declared in .metadata/com.nvidia.simready.root_usds.json. Address a prim inside one with USD's package-relative syntax:

<package>.usdz[<root>.usd]

Usage

from huggingface_hub import hf_hub_download, snapshot_download

# One file.
p = hf_hub_download(
    "imagine-io/PhysicalAI-Simready-Kitchen-Assets",
    "Models/appliances/Air_fryer_1.usdz",
    repo_type="dataset",
)

# Just the HDRIs.
snapshot_download(
    "imagine-io/PhysicalAI-Simready-Kitchen-Assets",
    repo_type="dataset",
    allow_patterns=["hdri/**"],
)

Files are also fetchable anonymously over plain HTTP at https://huggingface.co/datasets/imagine-io/PhysicalAI-Simready-Kitchen-Assets/resolve/main/<path>, which is what the reference arcs above rely on.

Resolving references

A USD runtime needs an HTTP-capable asset resolver to follow these URLs. NVIDIA Isaac Sim and Omniverse Kit resolve https:// asset paths. Stock OpenUSD with ArDefaultResolver (usdview, Blender's USD importer) does not, and will report the references as unresolved — for those consumers, download the assets locally or use a scene built with assets embedded.

Pin a commit sha rather than main if you need a fixed snapshot; main moves when assets are updated.