Hugging Face's logo

Datasets:
imagineio
/
PhysicalAI-SimReady-Homes

Tasks:
Robotics
Image Segmentation
Depth Estimation
Modalities:
3D
Image
Text
Formats:
parquet
Size:
1K - 10K
Tags:
simready
openusd
usd
usdz
isaac-sim
isaac-lab
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access to this dataset is granted for non-commercial research and evaluation. Please tell us who you are and what you plan to evaluate. Commercial training, fine-tuning, product development, or production use requires a separate written license from Imagine.

Log in or Sign Up to review the conditions and access this dataset content.

PhysicalAI SimReady Homes: Multi-Room Interiors

1,000 simulation-ready home interiors that load into Isaac Sim and are ready to train on — no rigging, no retopology, no physics authoring.

Every scene is a furnished multi-room home in OpenUSD with rigid bodies, mass and inertia, collision approximations, PhysX friction/restitution/density, articulated doors and drawers, per-object semantic labels, PBR materials, HDRI lighting and placed cameras — plus a machine-readable sidecar naming every prim by path, and an editable scene graph if you want to change the furniture rather than the pixels.

Built by Imagine.io. Companion to PhysicalAI SimReady Assets, the public object/material/HDRI library these scenes draw from.

Scenes 1,000
Objects per scene ~319 — furnished, collided, labelled
Distinct assets used 512
Floorplans 4 layouts (1br, 2br, 3bed, 4br), each furnished 250 ways
Total size 1.55 GB for the whole corpus (mean 1.55 MB per scene)
Lighting / surfaces 6 HDRI environments, 25 floor material packs
Units metres, Z-up, right-handed, kilogramsPerUnit=1
Format OpenUSD .usdz + metadata.json + scene_graph.json + floorplan.json + plan sheet

Why robotics teams use it

  • It is SimReady, not "a mesh you could simulate." Physics is authored per object class from a taxonomy — a chair is a dynamic rigid body with a convex-hull collider and wood friction; a countertop is static; a cabinet door is a joint with limits. You spawn a robot and the scene behaves. Nothing to clean up first.
  • Real interiors, not boxes on a plane. Multi-room homes with correct anthropometry: walkable circulation, doors that swing into rooms that have space for them, kitchens tiled with real cabinet runs and appliance clearances, bathrooms that actually seat a vanity, WC and shower. Layout comes from a space-planning engine, not from random scatter.
  • Every variation axis is a filterable column. Kitchen shape, sofa shape, TV mount, floor pack, wall tile, HDRI, sun azimuth, colour temperature, object histogram, room areas, footprint — all in the parquet index. Build a curriculum or a held-out split with a pandas query instead of eyeballing renders.
  • Scenes are under a megabyte. Geometry and textures resolve from the public assets repo by URL, so you can pull the whole corpus, or 20 scenes matching a spec, in seconds.
  • Reproducible and traceable. Each row carries code_rev, asset_index_sha256, seed and build flags. Bundles are content-named by a digest of their scene graph, so identical scenes are identical files.

What people build with it

Task What the data gives you
Indoor navigation / SLAM / mapping multi-room footprints, doorways, real clearances, floorplan.json as ground truth
Mobile manipulation articulated doors, cabinet doors and drawers with joints and limits; graspable props on reachable surfaces
Perception (segmentation, depth, pose) per-object semantic labels + physics_index prim paths → Isaac Replicator ground truth
Domain randomization / sim2real material packs, HDRI, sun angle and colour temperature as explicit axes, already varied across the corpus
VLA / embodied agents room-typed scene graph with poses, so instructions can be grounded ("go to the kitchen sink")
Layout & scene-synthesis research the authored scene_graph.json and source floorplan next to the built result

Quick start 

pip install huggingface_hub pandas pyarrow
huggingface-cli login                    # gated: request access first

python dataset_tools/download_scene.py --style scandinavian_3bed --seed 7 --out ./scenes --extract
python dataset_tools/validate_scene.py ./scenes/scandinavian_3bed_s7
<isaac>/python.sh dataset_tools/load_isaac.py ./scenes/scandinavian_3bed_s7/model.usdz

dataset_tools/ in this repo holds download_scene.py, extract_scene.py, validate_scene.py and load_isaac.py.

Two things to know before you load

1. A scene needs network access to open. The bundle holds the USD, not the geometry. Every mesh, PBR texture and the HDRI resolves by https:// URL from the public, ungated imagineio/PhysicalAI-SimReady-Assets at load time. That is why a scene is under a megabyte instead of about a gigabyte, and it means you need a resolver that speaks HTTPS — Isaac Sim's OmniUsdResolver does, the PyPI usd-core build does not (it opens the stage and shows an empty room). Access here is enough; the assets repo needs no separate grant.

If assets come up missing after a network blip, Omniverse caches the failed fetch. Clear %LOCALAPPDATA%\ov\cache (Windows) or ~/.cache/ov (Linux) before retrying.

2. A scene is one archive, and it nests. Download <scene_digest>.zip, unpack it, then open model.usdz — you cannot point a viewer at the download. There is no text-readable USD layer at any level; the root inside the package is a binary crate (model.usdc).

What's in each scene 

scenes/<scene_digest>.zip                ~1.55 MB
    model.usdz          the scene: geometry, physics, materials, lighting, cameras
    metadata.json       SimReady sidecar — rooms, objects, physics index, design choices
    topdown.png         plan sheet (imperial drafting sheet)
    scene_graph.json    editable room/object graph with poses and asset refs
    floorplan.json      the source floorplan the scene was built from

The archive is named by a digest of its scene graph, so identical scenes are identical files. layouts/<style>/ carries each floorplan's authored inputs once.

Use the hf_path / zip_url column, don't build the path. Earlier scenes are published at scenes/<style>/<scene_digest>.zip and both layouts are served, so each row's own path is the only correct one for that row.

Physics and SimReady, in detail

Every scene is authored for direct use in Isaac Sim:

  • Rigid bodies with mass and inertia, and static bodies for architecture and built-ins.
  • Collision approximations chosen per object class — convex hull, box proxy or SDF as the class warrants, rather than one blanket setting.
  • PhysX materials with real friction, restitution and density per surface (wood, tile, fabric, stone, metal).
  • Articulated openings — entry doors, cabinet doors and drawers as joints with limits and articulation roots, so a policy can actually open them.
  • Semantic labels per object, from a taxonomy that maps each class to its labels, body type, collider and physics preset.
  • Lighting and cameras — an HDRI dome plus ceiling fixtures, and per-room camera standpoints framed like an architectural photograph.

metadata.json carries a physics_index naming every rigid body, static body, collider, joint, articulation root, material, light and camera by prim path — so you can wire up Replicator, attach sensors or query the scene without walking the stage yourself.

The design block records the seeded per-room choices — the door style a kitchen's casework wears, the countertop stone, the sofa shape — at the point each was drawn. These are not derivable from the geometry, and they are mirrored into the index columns below.

Browse and filter the index

index/*.parquet is the index — and because nothing inside a bundle has its own URL, it is the only index. Every variation axis is a column, so you filter here and download only what you want. The preview column embeds a 256 px thumbnail, so the Data Studio viewer above shows a plan sheet for every row.

It is a directory of parquet shards, not one file: scenes are generated on several machines at once and each publishes its own rows as it uploads, so the index is always current rather than assembled afterwards. Every shard carries an identical schema, so the split loads as one table — read it through datasets, or glob it:

from datasets import load_dataset
df = load_dataset("imagineio/PhysicalAI-SimReady-Homes", split="train").to_pandas()

# or without `datasets`, straight from the parquet:
#   from huggingface_hub import snapshot_download
#   import pandas as pd, glob
#   d = snapshot_download("imagineio/PhysicalAI-SimReady-Homes", repo_type="dataset", allow_patterns="index/*.parquet")
#   df = pd.concat(pd.read_parquet(p) for p in glob.glob(f"{d}/index/*.parquet"))

# every 3-bedroom scene with an island kitchen and a wall-mounted TV
hits = df[(df.bedrooms == 3) & (df.interior_run == "island") & (df.tv_mount == "wall")]
print(hits[["style", "seed", "scene_id", "zip_url"]])

# a held-out split that shares no floorplan with training
train = df[df.style != "scandinavian_3bed"]
test  = df[df.style == "scandinavian_3bed"]

Drop the preview column if you only want to filter — it is ~34 KB of thumbnail per row and dominates the read.

Variation axes you can filter on

Group Columns
Identity scene_id (= scene_digest), style, seed, zip_url, hf_path
Plan shape room_count, bedrooms, bathrooms, room_types, room_areas_m2, floor_area_m2, footprint_w_m, footprint_d_m
Typology circulation, open_kitchen, has_courtyard, veranda_depth
Surfaces floor_pack_living, floor_pack_kitchen, floor_pack_bathroom, floor_pack_balcony, wall_pack_exterior, wall_pack_interior, ceiling_pack, wall_tile_pack, wall_tile_m, rug_packs, mat_packs
Kitchen design kitchen_shape, kitchen_run_count, cabinet_door_style, cabinet_handle_variant, countertop_finish, cook_mode, oven_housing, interior_run
Living design sofa_shape, tv_mount
Lighting hdri, lighting_mode, sun_azimuth, color_temp, light_count
Assets asset_ids, asset_names, assets_by_class, door_assets, window_assets, balcony_door_assets, object_variants
Composition object_count, referenced_count, placeholder_count, class_histogram
Physics rigid_body_count, static_body_count, collider_count, joint_count, articulation_root_count, material_count, camera_count
Quality dropped_required, dropped_bathroom_core, scale_up_factor, scale_up_capped, bath_openings_clamped
Provenance code_rev, asset_index_sha256, generated_at, remote_assets_remote, remote_assets_local, build_s, stage_timings_s

Metadata schema

metadata.json is simready_scene_metadata_v2: schema, schema_version, scene_id, units, coordinate_system, usd_export, packaging, variation_config, source_floorplan, functional_layout, dropped_furniture, design, rooms, objects, physics_index, build_stats.

Reproducibility

Each row carries code_rev, asset_index_sha256, seed and the build flags, which together name the exact inputs a scene came from.

One caveat, stated plainly: asset URLs point at resolve/main/ of the assets repo rather than a pinned commit. Scenes therefore track that repo's main branch — asset fixes reach already-published scenes for free, but a restructuring there would affect every scene here at once.

Known limits

We would rather you find these here than after a training run.

  • Walls are fixed per floorplan. A seed re-furnishes a home; it does not redraw it. Room polygons, door positions and the footprint are identical across every seed of a floorplan, so architectural diversity is 4 architectures. Split held-out sets by style, not by seed. Need more architectures? See below — we generate them on request.
  • Ceiling material is constant. The library ships exactly one ceiling pack, so every scene shares it.
  • Wall materials are thin — 7 packs across the corpus.
  • HDRI is keyed on the seed alone, from a pool of 6, so all floorplans at a given seed share their lighting.
  • Scenes are not SimReady-validated per scene in this release. The rules exist in the generator and pass on spot-checked scenes, but the corpus run did not gate on them.

License

CC BY-NC 4.0 — non-commercial research and evaluation. Commercial training, fine-tuning, product development or production use requires a separate written license from Imagine. See the access form above.

Citation 

@misc{imagineio_physicalai_simready_homes,
  title  = {PhysicalAI SimReady Homes: Multi-Room Interiors},
  author = {Imagine.io},
  year   = {2026},
  url    = {https://huggingface.co/datasets/imagineio/PhysicalAI-SimReady-Homes}
}

About Imagine.io — and getting a corpus shaped to your task

Imagine.io builds 3D content and simulation pipelines for retail, furniture and robotics. This corpus is a sample of what the pipeline produces: a text brief, a floor-plan image or a house style goes in, and furnished, physics-authored SimReady USD comes out — at whatever scale you need, with your own floorplans, your own product catalogue, or the room types and object classes your policy is failing on.

If that is useful to you, get in touch — more architectures, more material variety, per-scene validation and commercial licensing are all available.
Downloads last month
-
Total file size:
1.59 GB