nvidia/PhysicalAI-Robotics-NuRec

Dataset Description

The Physical AI NuRec dataset seeks to empower robotic researchers to build the next generation of physical AI based end-to-end robotic models.

This dataset contains reconstructed 3D scenes distributed as USD files and USDZ packages for use in NVIDIA Isaac Sim. Some datasets also include a mesh and occupancy map. The Mesh components are used for collision detection while the 3DGUT components provide realistic rendering. The asset can also be used with Isaac Sim Extensions like MobilityGen for AMR data generation.

Release Note · Asset Types · Quick Start · Dataset Details · Known Limitations

Dataset Owner(s)

NVIDIA

Release Note

Date	Tag / Branch	Isaac Sim	Changes
2026-10-09	main	6.1	Updated stereo assets; added the Bluefield iPhone manipulation scene.
2026-04-20	v0.1	5.1	Separate particle and volume exports; some scenes retain stage.usdz.
2026-03-10	v0.0	5.1	First tagged snapshot.
2025-10-29	—	—	Added hand-held Andoria, Livingroom and Wormhole scenes.
2025-10-07	—	—	Added Zurich fourth-floor iPhone scene.
2025-06-06	—	—	Added Nova Carter scenes.
2025-06-02	—	—	Added Zurich lounge scene.

Asset Types

PPISP (Physically Plausible Image Signal Processing) applies camera color correction. The current assets provide the following rendering variants:

Asset	Description
particle_sh_optimized.usdz	Particle-field asset with PPISP color correction baked into spherical-harmonic (SH) coefficients.
particle_spg-runtime.usdz	Particle-field asset with PPISP runtime controllers through a SensorProcessingGraph (SPG).
volume.usdz	Neural-volume asset with PPISP runtime controllers

Quick Start

Sign in to Hugging Face and accept the dataset access conditions, then install the CLI and authenticate once:

python3 -m pip install -U huggingface_hub
hf auth login

Download

Set ASSET_FILE to choose an asset and --revision to select a different version.

ASSET_FILE="particle_spg-runtime.usdz"  # or "volume.usdz" or "particle_sh_optimized.usdz"
hf download nvidia/PhysicalAI-Robotics-NuRec \
  --repo-type dataset --revision main \
  "nova_carter-galileo/${ASSET_FILE}" \
  "nova_carter-galileo/occupancy_map.yaml" \
  "nova_carter-galileo/occupancy_map.png" \
  --local-dir ./PhysicalAI-Robotics-NuRec

PPISP assets

For volume.usdz and particle_spg-runtime.usdz:

1. Launch Isaac Sim with SPG enabled:

./isaac-sim.sh \
  --/renderer/multiGpu/enabled=false \
  --/rtx/spg/enabled=true \
  --/omni/rtx/nre/compositing/disableNuRecPostProcessings=true \
  --/rtx/rtpt/gaussian/skipTonemapping/enabled=false \
  --enable omni.rtx.spg

2. Load the scene, then apply PPISP using the Script Editor:

Open the Script Editor and run the script below.

import re

import omni.kit.commands
import omni.usd
from omni.kit.viewport.utility import get_active_viewport
from pxr import Sdf, Usd, UsdGeom, UsdRender


DEFAULT_PERSPECTIVE_CAMERA = "/OmniverseKit_Persp"


def _copy_camera_to_default(stage, source_camera_path: Sdf.Path):
    source = UsdGeom.Camera.Get(stage, source_camera_path)

    if not source:
        raise RuntimeError(f"Camera does not exist: {source_camera_path}")

    target = UsdGeom.Camera.Define(stage, DEFAULT_PERSPECTIVE_CAMERA)
    source_prim = source.GetPrim()
    target_prim = target.GetPrim()

    # The *_ppisp cameras use these API schemas.  The default Kit camera does
    # not normally have them, so only copying regular UsdGeom.Camera fields is
    # insufficient when the stage is saved and reopened.
    api_schemas = source_prim.GetMetadata("apiSchemas")
    if api_schemas:
        target_prim.SetMetadata("apiSchemas", api_schemas)

    # Copy all authored camera attributes, including custom ppisp:* and
    # exposure:* attributes.  This also preserves attributes added by newer
    # camera extensions without maintaining a hard-coded list here.
    for src_attr in source_prim.GetAuthoredAttributes():
        name = src_attr.GetName()
        if name.startswith("xformOp:"):
            continue

        value = src_attr.Get()
        if value is None:
            continue

        dst_attr = target_prim.GetAttribute(name)
        if not dst_attr:
            dst_attr = target_prim.CreateAttribute(name, src_attr.GetTypeName())
        dst_attr.Set(value)

    # Explicit identity exposure.  These are the values used by the source
    # *_ppisp cameras in this asset and avoid Kit's implicit camera exposure
    # defaults changing the result after reopening the saved USDA.
    identity_exposure = {
        "exposure": 0.0,
        "exposure:fStop": 1.0,
        "exposure:iso": 0.0,
        "exposure:responsivity": 1.0,
        "exposure:time": 1.0,
        "omni:rtx:autoExposure:enabled": False,
    }
    for name, value in identity_exposure.items():
        target_prim.GetAttribute(name).Set(value)

    # Copy the source camera's world transform to the viewport perspective camera.
    source_xform = UsdGeom.Xformable(source_prim)
    target_xform = UsdGeom.Xformable(target_prim)

    world_transform = source_xform.ComputeLocalToWorldTransform(Usd.TimeCode.Default())

    target_xform.ClearXformOpOrder()
    target_xform.AddTransformOp().Set(world_transform)


def show_camera_with_sensor_processing():
    viewport = get_active_viewport()
    if viewport is None:
        raise RuntimeError("No active viewport")

    stage = viewport.usd_context.get_stage()
    if stage is None:
        raise RuntimeError("No USD stage is open")

    # Find an authored RenderProduct linked to an asset camera.
    original_product_path = None
    original_camera_path = None

    for prim in stage.Traverse():
        # Checking the type name as well as IsA() is important for products
        # authored by an asset layer that does not have the Render schema
        # registered in the current Kit process.
        if prim.GetTypeName() != "RenderProduct" and not prim.IsA(UsdRender.Product):
            continue

        product = UsdRender.Product(prim)
        camera_rel = prim.GetRelationship("camera")
        targets = camera_rel.GetTargets() if camera_rel else []

        if not targets:
            targets = product.GetCameraRel().GetForwardedTargets()

        for target in targets:
            if str(target) == DEFAULT_PERSPECTIVE_CAMERA:
                continue
            if UsdGeom.Camera.Get(stage, target):
                original_product_path = prim.GetPath()
                original_camera_path = target
                break

        if original_product_path:
            break

    if original_product_path is None:
        raise RuntimeError("No RenderProduct linked to an asset camera was found")

    # Duplicate the complete RenderProduct, preserving its SensorProcessingGraph.
    safe_name = re.sub(r"[^A-Za-z0-9_]", "_", original_camera_path.name)
    duplicate_path = Sdf.Path(f"/Render/Viewport_{safe_name}")

    if stage.GetPrimAtPath(duplicate_path):
        omni.kit.commands.execute("DeletePrims", paths=[duplicate_path])

    omni.kit.commands.execute(
        "CopyPrim",
        path_from=str(original_product_path),
        path_to=str(duplicate_path),
    )

    duplicate_product = UsdRender.Product(stage.GetPrimAtPath(duplicate_path))

    # Make the duplicate render product render through the default viewport camera.
    duplicate_product.GetCameraRel().SetTargets([Sdf.Path(DEFAULT_PERSPECTIVE_CAMERA)])

    # Put the original *_ppisp camera's pose and camera attributes on the
    # default perspective camera.
    _copy_camera_to_default(stage, original_camera_path)

    viewport.camera_path = DEFAULT_PERSPECTIVE_CAMERA
    viewport.render_product_path = str(duplicate_path)

    return str(duplicate_path), str(original_camera_path)


render_product_path, camera_path = show_camera_with_sensor_processing()
print("Source camera:", camera_path)
print("Viewport RenderProduct:", render_product_path)

Non-PPISP assets

For particle_sh_optimized.usdz, XGRIDS assets and mono-workflow assets, launch Isaac Sim and directly load the scene, no additional launch arguments are required.

Dataset Details

Type	Capture	Workflow	Views
Nova Carter stereo	Four calibrated stereo cameras; 360-degree coverage	Stereo reconstruction	Ground-robot / AMR heights
Hand-held stereo	Hand-held stereo camera	Stereo reconstruction	Hand-held viewpoints
Monocular	Camera or smartphone photos	Mono reconstruction; COLMAP + 3DGUT	Hand-held viewpoints
XGRIDS	XGRIDS capture	LCC Studio export and packaging	Hand-held viewpoints

Nova Carter scenes

Captured with Nova Carter and Isaac ROS Nova Recorder. Front, left, right and back cameras provide rectified images with a pinhole OpenCV camera model. Keep novel views near typical AMR heights.

Scene	DEMO
Galileo Lab aisles, shelves and boxes.	Training view	Novel view	Mesh	Depth
Cafe Open cafe with natural lighting.	Training view	Novel view	Mesh	Depth
Wormhole Conference room with chairs and indoor lighting.	Training view	Novel view	Mesh	Depth

Hand-held stereo scenes

Indoor captures made with a hand-held stereo camera.

Scene	Description	Rendered view
Endeavor Andoria	Andoria area in the NVIDIA Endeavor building.
Endeavor Livingroom	Living room in the NVIDIA Endeavor building.
Endeavor Wormhole	Wormhole conference room.
Endeavor Wormhole Table	Table variant of the Wormhole conference room.
Voyager Babyboom	Babyboom room in the NVIDIA Voyager building.

Monocular scenes

Scene	Capture and reconstruction
Bluefield iPhone	iPhone room capture for manipulation; dense COLMAP reconstruction and 3DGUT.
Zurich Lounge	Third-floor office entrance lounge; approximately 300 photos from a Nikon Z7 with a 14–30 mm f/4 lens at 14 mm, f/4.
Zurich Fourth Floor	Office entrance area; approximately 500 photos from an iPhone 13 Pro ultrawide camera (13 mm equivalent, f/1.8).

Both Zurich captures were resized and corrected for vignetting, then processed with COLMAP/GLOMAP for poses and 3DGRUT for training and export. They use sparse reconstruction rather than Bluefield's dense initialization.

XGRIDS scenes

Scene	Description
Wormhole	Exported through XGRIDS/LCC Studio, with an additional variant containing simulation objects. Captured with XGRIDS Portal Cam.

Download complete XGRIDS and Zurich scene folders to preserve referenced files, including Zurich .nurec dependencies.

Training Trajectory Poses (.tum files)

Where provided, training_trajectory_poses.tum (or trajectory_poses.tum) is a file that stores the camera trajectory data collected during the training phase. The .tum format follows the TUM RGB-D benchmark format and contains the following data per line:

Format: timestamp tx ty tz qx qy qz qw

Headers:

• timestamp: Unix timestamp in seconds (16 decimal precision)
• tx, ty, tz: Translation components in meters (x, y, z coordinates)
• qx, qy, qz, qw: Quaternion rotation components (x, y, z, w)

Known limitations

• Speckles during fast motion (Isaac Sim 6.1): Gaussian assets may display white speckles during fast camera or object movement when Gaussian tone mapping is skipped.

  • Workaround: Launch Isaac Sim with --/rtx/rtpt/gaussian/skipTonemapping/enabled=false to reduce these artifacts. SPG-runtime and volume assets with PPISP are unaffected because their setup already applies this setting.

• Black or gray patches from Gaussian sorting (Isaac Sim 6.1): Particle USDZ assets (including SH-optimized and SPG-runtime exports) may exhibit dark patches or speckles due to an RTX sorting issue.

  • Workaround: Change the Gaussian sorting mode to rayHitDistance. This reduces the observed patches, though it may introduce secondary rendering artifacts.

Additional resources

• Neural Volume Rendering
• Using Generated Assets and Sim-Ready Assets in Isaac Sim
• Isaac Sim NuRec Rendering Utilities
• Training COMPASS with NuRec scenes

License/Terms of Use

cc-by-4.0

Intended Usage

Open sourcing for Isaac Sim use cases as part of NVIDIA physical AI dataset initiative. The dataset will be downloadable via Hugging Face.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

For Release on NVIDIA Platforms Only

Please report security vulnerabilities or NVIDIA AI Concerns on NVIDIA Security page.