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 includes various 3DGUT in USD files that can be loaded in 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.

Dataset Owner(s)

Di Zeng, Chirag Majithia, Harel Omer, Sameer Chavan, Isaac Deutsch, Weihan Wang

Dataset Creation Date

06/06/2025 for Nova Carter datasets

06/02/2025 for Zurich lounge dataset

10/07/2025 for Zurich fourth floor iphone dataset

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.

Dataset Characterization

Data Collection Method

Nova Carter Datasets

The raw data is collected with Nova Carter and ISAAC ROS Nova Recorder, conducted through systematic planning in a controlled indoor environment with optimal lighting conditions. The setup consisted of:

• Four calibrated stereo cameras mounted on the Nova Carter robot platform
• Strategic camera placement (front, left, right, back) to achieve comprehensive 360-degree coverage
• Pinhole OpenCV camera model with post-capture image rectification
• Applied image rectification to raw camera feeds

This methodical approach to data collection was designed to maximize the quality and completeness of the dataset, particularly for applications requiring accurate scene reconstruction and novel view synthesis.

This raw data was passed through our NuRec pipeline to generate this dataset. The major difference of this dataset and other open-sourced 3DGS datasets is, we include corresponding mesh and occupancy maps in the USDZ file which are aligned with the 3DGS, this makes it Sim-Ready.

The Nova Carter datasets were created using the NuRec Stereo Camera Reconstruction Workflow, which processes stereo RGB camera data through an end-to-end pipeline involving Isaac ROS, cuSFM for pose estimation, FoundationStereo for depth estimation, nvblox for mesh generation, and 3DGURT for neural reconstruction.

Note that since data was captured by a ground robot for AMR applications, novel views should be in the range of common AMR heights.

Zurich Office Datasets

We provide two Zurich office datasets.

For "lounge", around 300 photos were taken with a Nikon Z7 camera and a 14-30mm f/4 lens, set at 14mm f/4.

For "fourth floor iphone", around 500 photos were taken with an iPhone 13 Pro, using the ultrawide lens (13mm equiv., f/1.8).

The images were resized and vignetting was removed, before being passed to COLMAP + GLOMAP for camera pose estimation.

The Zurich office datasets were created using the NuRec Mono Camera Reconstruction Workflow, which processes monocular camera data through COLMAP for pose estimation and 3DGURT for neural reconstruction.

The 3DGUT USD model was trained and exported using the open-source 3DGRUT repository.

Labeling Method

No labels are included in the dataset.

Dataset Details

Nova Carter Dataset

The assets are exported in USDZ format. This zip contains the mesh, rig trajectories, and other files required for rendering the scene in Isaac Sim 5.

Training Trajectory Poses (.tum files)

Each environment directory contains a training_trajectory_poses.tum file (or trajectory_poses.tum) 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)

Zurich Office Dataset

The COLMAP data is provided, with images zipped to reduce the file count.

The following files can be opened in Isaac Sim 5:

• "lounge": zh_lounge/usd/zh_lounge.usda
• "fourth floor iphone": zh_fourth_floor_iphone/usd/zh_4th.usda

Reproducing the Assets

The assets in this dataset can be reproduced using NVIDIA's NuRec reconstruction workflows. There are two workflows available depending on your camera setup:

Stereo Camera Workflow

For scenes captured with stereo RGB cameras (Nova Carter datasets), use the NuRec Stereo Camera Reconstruction Workflow. This workflow includes:

• ROS bag to cuSFM format conversion
• Pose estimation with cuSFM
• Depth estimation with FoundationStereo
• Mesh generation with nvblox
• Neural reconstruction with 3DGURT
• USDZ export for Isaac Sim

Mono Camera Workflow

For scenes captured with monocular cameras (Zurich office datasets), use the NuRec Mono Camera Reconstruction Workflow. This workflow includes:

• Pose estimation with COLMAP/GLOMAP
• Neural reconstruction with 3DGURT
• USDZ export for Isaac Sim

Both workflows produce Sim-Ready assets with aligned meshes, occupancy maps, and 3DGURT neural rendering components that can be directly loaded into Isaac Sim.

Dataset Quantification

We have provided the following different environments:

Environment Name	Description	Data Collection Platform	Workflow	Required Isaac Sim Version	MobilityGen Compatible
nova_carter-galileo	The Galileo lab in NVIDIA. This has aisles and shelves and boxes on shelves	Nova Carter	Stereo Camera	5.1	Yes
nova_carter-cafe	NVIDIA cafe, this is an open area with natural lighting conditions	Nova Carter	Stereo Camera	5.1	Yes
nova_carter-wormhole	A conference room in NVIDIA. This is a closed room with multiple chairs as objects and indoor room lighting conditions	Nova Carter	Stereo Camera	5.1	Yes
zh_lounge	Entrance lounge area of the NVIDIA Switzerland office, 3rd floor	Hand-held camera	Mono Camera	5.1	No
zh_fourth_floor_iphone	Entrance area of the NV CH office, 4th floor	Hand-held smartphone	Mono Camera	5.1	No

Demo Videos

Galileo Lab

Training View:

Novel View:

Mesh quality:

Depth:

Cafe

Training View:

Novel View:

Mesh quality:

Depth:

Wormhole

Training View:

Novel View:

Mesh quality:

Depth:

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.

How to Load Asset in Isaac Sim

For more detailed example of loading the NuRec simulation can refer this example.

For more details on how to add Sim-Ready Assets into NuRec scene can refer this example.

Load the galileo scene

1. Select File > Open.

2. Enter the asset path URL under File name:

   ex: "/home/workspaces/user/PhysicalAI-Robotics-NuRec/nova_carter-galileo/stage.usdz"

3. Click Open File.