Open Source RoboticsRobots, software & learning

Newton Physics Engine

Software

Linux Foundation GPU-accelerated, differentiable robotics physics engine built on NVIDIA Warp with MuJoCo Warp and OpenUSD support, licensed under Apache-2.0 (code) and CC-BY-4.0 (docs). [2] [3]

Installation & first result →Official websiteSource codeGetting started

For
Robotics simulation researchers; RL and sim-to-real engineers; Physics engine and differentiable-control developers [2] [4]
Hardware
NVIDIA Turing+ GPUs (compute capability 7.5+, driver R580+ for Warp 1.18+), CPU fallback (including macOS) [2]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
Ecosystem
MuJoCo

Requirements · Research details and dates

On this page

Overview

Newton is a Linux Foundation GPU-accelerated, differentiable physics engine for robotics built on NVIDIA Warp, initiated by Disney Research, Google DeepMind, and NVIDIA. It generalizes Warp's deprecated warp.sim module, uses MuJoCo Warp as its primary articulated-body backend, and adds native OpenUSD and URDF workflows. GitHub release metadata records v1.8.0 on 2026-11-11, while the README and pyproject.toml inspected in this review come from main commit ab4af1502ab9340bb5946665311da0d43795246c (checked 2026-11-21), which declares version 1.9.0.dev0. Because the v1.8.0 manifest was not separately inspected, attribute the requirements and optional extras below to that main checkout and check the v1.8.0 release metadata before assuming identical pins on PyPI. [1] [2] [4]

Is this for you?

Workflows
  • Install modular extras ([sim], [importers], [examples]) via pip or uv
  • Load URDF or OpenUSD scenes and simulate on GPU via MuJoCo Warp
  • Record and replay trajectories with built-in viewers
  • Run ONNX neural actuators / RL policy examples

[2] [4]

Known limitations
  • CUDA acceleration with Warp 1.18+ requires an NVIDIA Turing GPU (compute capability 7.5+) and driver R580+ (older R545–R579 drivers require pinning warp-lang==1.17.0); macOS runs on CPU only.
  • Base pip install newton only installs warp-lang>=1.17.0; MuJoCo Warp (newton[sim]), USD/mesh importers (newton[importers]), and viewer/examples (newton[examples]) are separate optional extras with platform/Python constraints (e.g., warp-optix is restricted to Python 3.12 on Linux/Windows x86_64).
  • Simulation throughput, differentiability convergence, and third-party community solvers were not benchmarked in this review.

[2] [4]

What you need

Operating system
  • Linux (x86-64, aarch64)
  • Windows (x86-64)
  • macOS (CPU only)

[2]

Languages / runtimes
  • Python (with NVIDIA Warp GPU kernels)

[2] [4]

Dependencies
  • Python >=3.11
  • warp-lang>=1.17.0
  • Optional [sim]: mujoco-warp~=4.0.1, mujoco-core~=4.0.1
  • Optional [importers]: usd-core, newton-usd-schemas>=0.7.0, trimesh, coacd, resolve-robotics-uri-py
  • Optional [onnx]: warp-nn[onnx]~=0.4.0

[4]

Host computer

NVIDIA Turing GPU (compute capability 7.5+) with driver R580+ for CUDA acceleration with Warp 1.18+ (or warp-lang==1.17.0 on R545–R579); no local CUDA Toolkit required; macOS runs on CPU only.

[2]

Compute for the workload
  • NVIDIA Turing GPU (compute capability 7.5+) and driver R580+ for Warp 1.18+ CUDA wheels (or warp-lang==1.17.0 on R545–R579)
  • CPU-only execution supported on macOS

[2]

Skills
  • Python 3.11+
  • NVIDIA Warp / GPU simulation
  • MuJoCo / OpenUSD / URDF scene setup
  • uv or pip packaging

Inferred from README quickstart, uv source workflow, and pyproject.toml extras.

[2] [4] · Editorial assessment

Installation & first result

Open the upstream getting-started guide [2] · Project documentation [1] [2]

  • Verify Python 3.11+ and (for CUDA) an NVIDIA Turing+ GPU with driver R580+ (or pin warp-lang==1.17.0 on R545–R579 drivers; no local CUDA Toolkit required).
  • Install Newton with examples via pip install "newton[examples]" (or select minimal extras such as newton[sim] and newton[importers]).
  • Inspect example specifications without executing them via python -m newton.examples --describe robot_omniwheel, or run python -m newton.examples basic_urdf.

[2] [4]

Installation

pip install "newton[examples]" (or uv run --extra examples from a source checkout); base package is pip install newton with modular extras [sim], [importers], [onnx], [optix], [remesh], and [examples].

[2] [4]

Hardware, models & data

Supported hardware
  • NVIDIA Turing+ GPUs (compute capability 7.5+, driver R580+ for Warp 1.18+)
  • CPU fallback (including macOS)

[2]

Interfaces
  • Python API (newton)
  • CLI example runner (python -m newton.examples)
  • OpenGL viewer (pyglet / imgui_bundle) and recording/replay viewer
  • URDF and OpenUSD scene importers

[2] [4]

Inputs and outputs

Loads URDF, OpenUSD, and mesh assets into GPU-accelerated Warp/MuJoCo Warp simulation models; outputs physics state trajectories, OpenGL/recording viewer streams, and binary CBOR recordings.

[2] [4]

File formats
  • OpenUSD (usd-core, newton-usd-schemas)
  • URDF
  • Mesh formats (trimesh, meshio, COLLADA .dae)
  • CBOR binary recordings (.bin)
  • ONNX policy models (via [onnx] extra)

[2] [4]

Algorithms and models
  • MuJoCo Warp articulated-body physics backend
  • NVIDIA Warp differentiable simulation kernels
  • Optional ONNX policy inference (warp-nn[onnx])

[2] [4]

Evaluation

Built-in example runner supports inspecting descriptions and success criteria via python -m newton.examples --describe <name>; not executed in this review.

[2]

Project notes

GPU driver floors and OS limits

Newton requires Python 3.11+ and runs on Linux (x86-64, aarch64), Windows (x86-64), and macOS (CPU only). No local CUDA Toolkit installation is required, but CUDA acceleration has an important driver threshold: starting with Warp 1.18, standard Warp wheels require an NVIDIA Turing GPU (compute capability 7.5+) and driver R580+. On older supported GPUs or hosts running drivers R545–R579, upstream instructs pinning warp-lang==1.17.0. [2]

Core package vs. optional extras in pyproject.toml

A bare pip install newton only pulls warp-lang>=1.17.0. Most robotics workflows require opting into specific extras declared in pyproject.toml: [4]

  • [sim]: pulls mujoco-warp~=4.0.1 and mujoco-core~=4.0.1 (excluded on Windows ARM64).
  • [importers]: pulls usd-core / usd-exchange (with Python-version and aarch64 wheel markers), newton-usd-schemas>=0.7.0, trimesh, coacd, scipy, and resolve-robotics-uri-py for local-file, package, and model robotics URIs.
  • [onnx]: pulls warp-nn[onnx]~=0.4.0 for neural actuators and RL policy examples.
  • [optix]: pulls warp-optix~=0.3.0 for hardware-accelerated camera traversal, currently restricted to Python 3.12 on Linux/Windows x86_64.
  • [examples]: bundles [sim], [importers], [onnx], pyglet, imgui_bundle, GitPython (for newton.utils.download_asset()), and cbor2 binary recordings.

Every bundled example embeds its description and success criteria, which you can inspect before executing anything via python -m newton.examples --describe robot_omniwheel (or uv run --extra examples -m newton.examples ... from a source checkout). [2]

What we checked

Primary-source repository metadata, README, LICENSE.md, and pyproject.toml inspection only. No Python environments were created, no GPU kernels were compiled, and no simulation examples were run.

Updates

Upstream last updated
Nov 21, 2026 [1]
Latest release
Nov 11, 2026 [1]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
newton-physics/newton@ab4af1502ab9340bb5946665311da0d43795246c · software repository

Source files and licenses

Flagship Linux Foundation GPU-accelerated robotics physics engine with verified Apache-2.0 code and CC-BY-4.0 documentation licensing.

LayerAvailability / license
softwareavailable · Apache-2.0
Code licensed under Apache-2.0; additional third-party license texts are bundled in newton/licenses/. [2] [3] [4]
documentationavailable · CC-BY-4.0
README explicitly declares Documentation is licensed under CC-BY-4.0. [2]

Physics simulation engine; no pretrained model weights distributed as core artifacts.

Physics simulation engine; no standalone datasets distributed.

Apache-2.0 code and CC-BY-4.0 documentation licensing

Newton separates its code and documentation rights: source code is licensed under Apache-2.0 (LICENSE.md), documentation is designated CC-BY-4.0 in the README, and bundled third-party license texts are collected under newton/licenses/. Under the supplied Apache-2.0 text, redistributions must include a copy of the license, carry prominent notices on modified files, retain applicable copyright, patent, trademark, and attribution notices in source-form distributions of derivative works (excluding notices that do not pertain to any part of the derivative works), and include applicable NOTICE attributions when distributing derivative works if the original includes a NOTICE file. The full CC-BY-4.0 legal text was not included or reviewed in this packet, so consult it directly before redistributing documentation. External community solvers tagged with newton-physics on GitHub are explicitly unendorsed and maintain their own licenses and support levels. [2] [3] [4]

No Python packages were installed, no Warp GPU kernels were compiled, and no simulation examples were executed in this review. The next step is verifying your host NVIDIA driver version (R580+ for Warp 1.18+, or pinning warp-lang==1.17.0 on R545–R579 drivers or older supported GPUs), installing newton[examples] in a Python 3.11+ environment (noting that compatibility between released PyPI packages and the inspected 1.9.0.dev0 development manifest has not been established here), and running python -m newton.examples basic_urdf to confirm local rendering and backend execution. [2] [4]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history