README.md

---
license: cc-by-nc-4.0
pretty_name: PhysicalAI SimReady Assets
size_categories:
  - 1K<n<10K
tags:
  - simready
  - openusd
  - usd
  - usdz
  - isaac-sim
  - omniverse
  - robotics
  - embodied-ai
  - synthetic-data
  - 3d-assets
  - pbr-materials
  - hdri
  - digital-twin
---

# PhysicalAI SimReady Assets

**The constant half of a simulation-ready home.** 764 SimReady object packages,
98 PBR surface packs and 13 environment maps — every asset needed to
furnish, finish and light a synthetic interior, published once so scenes never have to carry
their own copy.

A packaged SimReady home scene is typically **500 MB to 1.2 GB**, and ~99% of that is these
assets, embedded again in every single scene. Reference this dataset by URL instead and the same
scene ships as **a few megabytes** — the geometry that is genuinely unique to it, plus links.

Measured on four apartment scenes, embedding versus linking the identical content:

| scene | assets embedded | assets by URL | |
|---|---:|---:|---:|
| 1-bedroom | 502.0 MB | **3.42 MB** | 147× |
| 2-bedroom | 664.7 MB | **3.84 MB** | 173× |
| 3-bedroom | 700.4 MB | **3.13 MB** | 224× |
| 4-bedroom | 1227.2 MB | **3.78 MB** | 325× |
| **total** | **3,094.3 MB** | **14.17 MB** | **218×** |

```
4-bedroom, assets embedded  ████████████████████████████████████████  1227.2 MB
4-bedroom, assets by URL    ▏                                            3.78 MB
```

Bigger scenes shrink *more*, because their extra size was almost entirely duplicated assets.
The scenes are otherwise identical: same geometry, same poses, same physics — a remote build
reproduces the embedded build's physics index exactly (117 referenced assets, 374 rigid bodies,
1,057 colliders).

Everything here is *SimReady*, not merely a mesh: each object carries collision geometry, mass
and inertia, physics materials, semantic labels, and — where the object articulates — working
joints. They are made to be simulated, not just rendered.

---

## What's inside

| Path | Contents | Files | Size |
|---|---|---:|---:|
| `simready_assets/` | SimReady object packages (USD + textures + MDL + physics) | 11,841 | 6,529 MB |
| `materials/` | PBR surface packs (floor, wall, ceiling, cabinet, countertop, tile, rug, carpet) | 290 | 294 MB |
| `hdri/` | Equirectangular environment maps (day / evening / night pools) | 13 | 83 MB |
| `assets/` | In-repo part kits (balcony railings) | 42 | 4 MB |

**12,187 files, 6.9 GB total.**

### Objects — `simready_assets/`

764 packages spanning **76 semantic classes**; 489 carry
validated real-world dimensions in the catalog index. Each is an NVIDIA SimReady–format package,
extracted intact:

```
simready_assets/132/
├── oven.usd                                      # root layer (USD crate)
├── OmniPBR.mdl                                   # material
├── SubUSDs/
│   └── textures/oven_display_diffuse.jpg         # …and siblings
├── gltf/pbr.mdl
├── .metadata/
│   ├── com.nvidia.simready.root_usds.json        # which layer is the root
│   └── com.nvidia.simready.packaging.bom.json    # vendor sha256 + blake3
└── com.nvidia.simready.packaging.json            # owner + license
```

<details>
<summary><b>All 76 object classes</b></summary>

`appliance.cooktop`, `appliance.dishwasher`, `appliance.fridge`, `appliance.hood`, `appliance.microwave`, `appliance.oven`, `appliance.range`, `balcony_door`, `bathtub`, `bed`, `cabinet_base`, `cabinet_tall`, `cabinet_wall`, `cabinet_wall_microwave`, `ceiling_light`, `chair`, `clutter.appliance`, `clutter.book`, `clutter.bottle`, `clutter.cookware`, `clutter.cup`, `clutter.food`, `clutter.frame`, `clutter.lamp`, `clutter.mirror`, `clutter.plant`, `clutter.plate`, `clutter.small`, `clutter.toiletry`, `clutter.utensil`, `coffee_table`, `door`, `door_double`, `dresser`, `dustbin`, `faucet`, `fireplace`, `floor_lamp`, `floor_mirror`, `mirror`, `nightstand`, `ottoman`, `outdoor_chair`, `outdoor_table`, `plant`, `rug`, `shelf`, `shower`, `sink`, `socket`, `sofa`, `sofa_mod_armless`, `sofa_mod_chaise_left`, `sofa_mod_chaise_right`, `sofa_mod_corner`, `sofa_mod_end_left`, `sofa_mod_end_right`, `sofa_mod_loveseat_armless`, `sofa_mod_loveseat_left`, `sofa_mod_loveseat_right`, `sofa_mod_ottoman`, `stool`, `table`, `television`, `television_wall`, `toilet`, `toilet_paper`, `towel_bar`, `tv_unit`, `utensil_hanger`, `vanity`, `wall_art`, `wall_clock`, `wall_light`, `wardrobe`, `window`

</details>

Appliances, seating, casework, bathroom and kitchen fixtures, lighting, wall decor, plants and
small clutter — plus a modular sofa system (`sofa_mod_*`) whose corner, chaise, armless and
loveseat pieces compose into arbitrary sectionals.

### Surfaces — `materials/`

98 texture packs across 8 categories:
`cabinet`, `carpet`, `ceiling`, `countertop`, `floor`, `rugs`, `wall`, `wall_tile`.

Packs follow the AmbientCG naming convention (`*_Color`, `*_NormalGL` / `*_NormalDX`,
`*_Roughness`, `*_AmbientOcclusion`), so a loader can pick maps by suffix. **The folder name
encodes the physical tile size** — `floor/2_1.69_meter/` tiles at 1.69 m, which is what makes
correct real-world UV scaling possible without per-pack metadata.

> One exception worth knowing: under `wall_tile/`, the `_N_meter` number is *repeats across the
> widest band* (a `cube_project` count), not a physical size.

All maps are power-of-two. That is not cosmetic — non-power-of-two textures render blue or
missing under Isaac Sim's RTX renderer while looking fine in Blender.

### Lighting — `hdri/`

13 equirectangular environment maps in **day / evening / night** pools, at `1k` and `2k`.
Intended as a `UsdLuxDomeLight` texture. Filenames are stable identities across resolutions, so a
scene that picks an HDRI by name renders consistently whichever variant you resolve.

---

## Two ways to use it

### 1. Download it, build against it locally

```bash
pip install huggingface_hub

# everything (6.9 GB)
hf download imagineio/PhysicalAI-SimReady-Assets --repo-type dataset --local-dir ./simready-assets

# or just what you need
hf download imagineio/PhysicalAI-SimReady-Assets --repo-type dataset --local-dir ./simready-assets \
    --include "simready_assets/132/*" "materials/floor/*" "hdri/2k/*"
```

```python
from pxr import Usd, UsdGeom

stage = Usd.Stage.CreateNew("scene.usda")
oven = UsdGeom.Xform.Define(stage, "/World/Oven")
oven.GetPrim().GetReferences().AddReference("./simready-assets/simready_assets/132/oven.usd")
stage.GetRootLayer().Save()
```

### 2. Reference it by URL — no download

Every file is served at a stable, public URL:

```
https://huggingface.co/datasets/imagineio/PhysicalAI-SimReady-Assets/resolve/main/<path>
```

```python
BASE = "https://huggingface.co/datasets/imagineio/PhysicalAI-SimReady-Assets/resolve/main"

oven.GetPrim().GetReferences().AddReference(f"{BASE}/simready_assets/132/oven.usd")
dome.CreateTextureFileAttr(f"{BASE}/hdri/2k/DayEnvironmentHDRI022_4K_HDR.exr")
```

> ### ⚠️ Your USD runtime must resolve `https://`
>
> This is the one thing to check before committing to URL references.
>
> | Runtime | Resolver | `https://` refs |
> |---|---|---|
> | **Isaac Sim / Omniverse Kit** | `OmniUsdResolver` (primary) | ✅ resolved natively |
> | **`usd-core` from PyPI** | `ArDefaultResolver` | ❌ not a registered URI scheme |
>
> Stock `usd-core` does not merely fail to fetch — it treats the URL as a file path and collapses
> `https://` to `https:/` before giving up. Pipeline stages that must run offline (validation,
> CI, unit tests) need the downloaded copy from option 1.
>
> Note also that a `.usdz` containing external references is **outside the USDZ specification**,
> which requires a package to be self-contained. Kit will open one; other tools may refuse. For a
> URL-referencing scene, prefer a plain `.usda` / `.usdc`.

**Pin a revision for reproducibility.** `resolve/main` follows the branch. Substituting a
commit SHA freezes the bytes, which is what you want for a dataset artifact whose geometry was
measured against a specific version of these assets:

```
https://huggingface.co/datasets/imagineio/PhysicalAI-SimReady-Assets/resolve/main/…            # follows main
.../resolve/<commit-sha>/…   # frozen
```

---

## Provenance and processing

Published by **imagine.io** from its SimReady asset catalog. The set is bounded and reproducible
rather than open-ended: `pinned_ids.json` at the repo root lists exactly which object ids belong
to it, so "did the set change?" is a question with an answer.

Each package retains the vendor's own `sha256` + `blake3` integrity BOM under `.metadata/`, so you
can verify any file against what was published.

Two processing passes have been applied, both deliberate:

- **Textures are capped** at 2048 px albedo / 1024 px response maps. The uncapped set contained
  8192×8192 normal maps on an egg boiler and two 81 MB diffuse PNGs on a single dining table —
  5.89 GB of resolution that no renderer benefits from at furniture scale.
- **Inverted collider windings are repaired.** Inside-out SDF collider meshes make PhysX compute
  negative mass, which fails at `attachShape` rather than at load.

Pre-processing backups (`*.srfp_orig`) are **not** published — they are the bytes that were
deliberately replaced.

---

## License

**CC BY-NC 4.0** — [Creative Commons Attribution-NonCommercial 4.0 International](https://creativecommons.org/licenses/by-nc/4.0/).

Free to use, modify and redistribute for **non-commercial** purposes with attribution. Every
object package self-declares this license in its own
`com.nvidia.simready.packaging.json`, and scenes built from these assets carry it forward in their
packaging manifest.

For commercial licensing, contact [imagine.io](https://imagine.io).

## Citation

```bibtex
@misc{imagineio_physicalai_simready_assets,
  title        = {PhysicalAI SimReady Assets},
  author       = {imagine.io},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/imagineio/PhysicalAI-SimReady-Assets}},
  note         = {CC BY-NC 4.0}
}
```

---

<sub>Generated by `scripts/hf_assets.py card` from the live asset roots — counts and sizes reflect
revision `main`.</sub>