Install Newton via conda-forge using pixi! 🎉

[diegoferigo]

We're all excited about the first stable release of Newton, arriving right on time for GTC! 🚀

Over the past couple of weeks, the conda-forge community has been working hard to package this new warp-based simulation stack.

Tip

I'm happy to announce that you can now install Newton (unofficially, for now) directly into a conda environment.

While package managers like uv are fantastic and blazing-fast for purely Python ecosystems, using conda-forge alongside pixi brings a few specific advantages for a complex, GPU-accelerated stack:

1. Unified dependency graph

PyPI wheels often bundle their own C++ libraries to remain self-contained, which can lead to bloat or subtle ABI clashes. Conda-forge natively resolves shared C++ libraries and rendering engines in a single unified graph.

Note

For example: On PyPI, usd-exchange vendors its own copy of usd-core (OpenUSD). With conda-forge, you can cleanly install both usd-core and usd-exchange in the exact same environment without conflicting or duplicating binaries, opening the possibility to have Newton installed alongside the *-usd-converter packages.

2. Zero host system requirements

There might be cases where installing system-level packages via apt is required. With conda-forge, every system dependency is fetched and isolated within the environment. There is no requirement for anything installed in the host on any of the supported platforms.

3. Comprehensive lockfiles

Both tools provide excellent lockfiles, but pixi locks your Python packages alongside all those lower-level dependencies. This guarantees true cross-platform reproducibility without ever requiring OS-level package managers like apt or brew.

Quickstart guide

If you'd like to test this setup, here is how you can get started quickly using pixi for convenience:

# Install pixi (in ~/.pixi).
# https://pixi.prefix.dev/latest/installation/
curl -fsSL https://pixi.sh/install.sh | sh
source ~/.bashrc

# Create a new environment.
mkdir newton_env && cd newton_env
pixi init

# Enforce CUDA driver version (not mandatory), options are 12 and 13.
# Requires only a compatible driver on the host, no CUDA required there.
pixi workspace system-requirements add cuda 13

# Enable all supported platforms.
# Already works on Jetson Thor with CUDA 13 (and on the Orin after upcoming JP7 release).
pixi workspace platform add linux-64
pixi workspace platform add osx-arm64
pixi workspace platform add linux-aarch64

# Platform supported soon.
# pixi workspace platform add win-64

# Add newton (and few other optional deps) and resolve the environment for all enabled platforms.
pixi add newton-all pytorch usd-exchange

# Run demos (on GPU if available).
pixi run python -m newton.examples robot_policy
pixi run python -m newton.examples mpm_anymal
pixi run python -m newton.examples nut_bolt_sdf
pixi run python -m newton.examples ik_cube_stacking

Kudos to the development team for delivering a project of such high quality! 🎉

Warning

As this is the first packaged version of Newton on conda-forge, please report any problems if you find them.

cc @eric-heiden @adenzler-nvidia @shi-eric @preist-nvidia @mmacklin, and the whole upstream development team.
cc @traversaro @ruben-arts @wolfv on the pixi / conda-forge side.

[ruben-arts]

Short video

unitree_newton_sim.mov

Amazing work you all! I had to try this quickly on my Macbook.
This one command installed the environment, cloned all data and compiled the simulation. What a crazy experience! 2 minutes in and it was running!

pixi exec -s newton-all -s pytorch -s usd-exchange python -m newton.examples robot_policy

[diegoferigo-rai]

Thanks for the feedback, @ruben-arts!

For the records, I've also addressed the last remaining blockers for linux-aarch64. Newton can now run on ARM machines and, with CUDA 13, on Jetson boards such as Thor, which I tested successfully this morning 🤖

[ruben-arts]

That's so cool! Is only CUDA >=13 supported?

[diegoferigo]

Yes, CUDA 13 only. Tegra support on conda-forge is still very limited, and warp-lang does not support it yet.

[flferretti]

Crossref for warp-lang conda-forge/warp-lang-feedstock#98 :)

[preist-nvidia]

Awesome, thank you @diegoferigo !