Installation#

This page describes how to install nvblox_torch (python) and nvblox (c++).

Supported Platforms#

The following platforms are supported:

x86 + dGPU

Jetson (ARM64)

nvblox_torch

✅

❌

nvblox

✅

✅

We support the systems with the following configurations:

  • x86 + discrete GPU

    • Ubuntu 20.04, 22.04, 24.04

    • CUDA 11.4 - 12.8

  • Jetson (ARM64)

    • (ARM64) Jetpack 5, 6

A minimum NVIDIA driver version is imposed by the version of CUDA you have installed. See the support table here to find the minimum driver version for your platform.

nvblox_torch#

There are two ways to install nvblox_torch:

  1. Install nvblox_torch via pip

  2. Install nvblox_torch from Source (in Docker)

pip is the preferred way to install nvblox_torch on Supported Platforms. Source installation is only recommended for developers who need to modify nvblox_torch or for platforms that are not supported via pip.

Install nvblox_torch via pip#

To install nvblox_torch via pip on a supported platform, run the following commands:

sudo apt-get install python3-pip libglib2.0-0 libgl1 # Open3D dependencies
pip3 install https://github.com/nvidia-isaac/nvblox/releases/download/v0.0.8/nvblox_torch-0.0.8rc5+cu12ubuntu24-863-py3-none-linux_x86_64.whl

(Optional) You can verify the installation by running our tests:

cd $(python3 -c "import site; print(site.getsitepackages()[0])")/nvblox_torch
pytest -s

You’re all set! You can now run the 3D Reconstruction example.

Install nvblox_torch from Source (in Docker)#

The source installation is recommended for developers who need to modify nvblox_torch or for platforms that are not supported via pip. We provide a docker image for building and developing inside.

First clone the repository:

git clone git@github.com:nvidia-isaac/nvblox.git

Then build and run the docker container:

cd nvblox
./docker/run_docker.sh

To build the c++ library run

mkdir -p /workspaces/nvblox/build
cd /workspaces/nvblox/build
cmake ..
make -j${nproc}

To install nvblox_torch in development/editable mode run

cd /workspaces/nvblox/nvblox_torch
pip3 install -e .

(Optional) You can verify the installation by running our tests:

cd /workspaces/nvblox/nvblox_torch
pytest -s

You’re all set! You can now 3D Reconstruction.

nvblox#

We support two installation methods for building the nvblox c++ library:

  1. Install nvblox from Source (in Docker) (recommended)

  2. Install nvblox from Source (Outside Docker)

After installing either way, you’re ready to Run an Example.

Install nvblox from Source (in Docker)#

The steps to build nvblox in the development container are the same as the instructions in Install nvblox_torch from Source (in Docker). nvblox is built inside our development container as the first part of installing nvblox_torch.

One difference is that on Jetson platforms we need to disable building of the pytorch wrapper, which is (currently) only supported on x86 platforms. Also note that the Jetson docker build only supports Jetpack 6.2. The modified (jetson) and unmodified (x86) building commands are:

mkdir -p /workspaces/nvblox/build
cd /workspaces/nvblox/build
cmake ..
make -j${nproc}

(Optional) To confirm building was a success, run the tests:

cd nvblox
ctest

You’re now ready to Run an Example.

Install nvblox from Source (Outside Docker)#

These instructions describe how to install the nvblox core library from source, outside of our development container.

Note

We recommend using the Install nvblox from Source (in Docker) as it will handle all the dependencies for you. The docker image sets up a controlled environment in which we know things work. While we’ve tested the following instructions on many systems (see Supported Platforms), results may vary.

To start, install our dependencies

sudo apt-get update && sudo apt-get install cmake git jq gnupg apt-utils software-properties-common build-essential sudo python3-pip wget sudo git python3-dev git-lfs

Note that for Ubuntu 20.04, we need to install a more recent version of cmake than is available in the default repositories. We provide a script to add the relevant repositories and install a more recent version in: docker/install_cmake.sh. Note that running this script will replace any previously installed version of cmake.

Now follow the instructions in Install nvblox from Source (in Docker) to build the code and run the tests.

You’re now ready to Run an Example.

Advanced Build Options#

This section details build options for advanced nvblox users.

Modifying maximum feature size#

The library supports integrating generic image features into the reconstructed voxel map. The maximum supported length of image feature vectors is a compile-time constant which defaults to 128. To change the default, call cmake with the following flag:

cmake -DNVBLOX_FEATURE_ARRAY_NUM_ELEMENTS=XYZ ..

Note that increasing this number will approximately linearly increase memory usage for applications using deep feature mapping.

Building for Post-CXX11 ABI#

The library is built with the pre-cxx11 ABI by default in order to maintain compatibility with manylinux201X wheels. To build with the post cxx11 ABI, call cmake with the following flag:

cmake -DPRE_CXX11_ABI_LINKABLE=OFF ..

Disabling pytorch wrapper#

If you don’t need the pytorch wrapper, or you’re on a system without pytorch installed, you can disable it by calling cmake with the following flag:

cmake -DBUILD_PYTORCH_WRAPPER=0 ..

Other docker containers#

We build and test in the following docker images, so if you would like to install in a docker, and don’t want to use our development docker, these are guaranteed to work.

  • nvcr.io/nvidia/cuda:12.8.0-devel-ubuntu24.04

  • nvcr.io/nvidia/cuda:12.6.1-devel-ubuntu22.04

  • nvcr.io/nvidia/cuda:11.8.0-devel-ubuntu22.04

Build a Redistributable Library#

By default, the nvblox library only builds for the Compute Capability (CC) of the GPU in the machine it’s being built on. Sometimes it is desirable to build a library that can be used across multiple machines that contain GPUs with different architectures. We, for example, build nvblox for several architectures for packaging into our pip package nvblox_torch, such that it can be used on a variety of machines.

To build binaries that can be used across multiple machines like this, you can use the CMAKE_CUDA_ARCHITECTURE flag and set it to a semicolon-separated list of architectures to support.

For example, to build for Compute Capability (CC) 7.2 and 7.5, you would run:

cmake .. -DCMAKE_CUDA_ARCHITECTURES=75;72