The gap between robotics research and real-world deployment has always been wide. Simulations that work beautifully in the lab often fall apart when transferred to physical hardware, and the tools to bridge that divide have historically been expensive, closed, or both. Newton, a newly open-sourced GPU-accelerated physics simulation engine built on NVIDIA Warp, aims to change that calculus — and it is already turning heads on GitHub.
Newton is not a toy project. It is a Linux Foundation initiative backed by Disney Research, Google DeepMind, and NVIDIA itself. With over 4,000 GitHub stars in a matter of days, the project is signaling that the robotics and simulation community is hungry for a modern, open, GPU-native physics engine that prioritizes differentiable simulation and enterprise-grade extensibility.
Built on the Shoulders of Giants
Newton extends and generalizes NVIDIA’s Warp framework, specifically the now-deprecated warp.sim module. It integrates MuJoCo Warp — developed by Google DeepMind — as its primary backend. The result is a physics simulation stack that combines Warp’s GPU-accelerated primitives with MuJoCo’s battle-tested contact and constraint handling.
The project emphasizes four core design principles: GPU-based computation (everything runs on the GPU by default), OpenUSD support (industry-standard scene description for seamless integration with film, gaming, and industrial pipelines), differentiability (physics gradients that enable reinforcement learning and model-based control), and user-defined extensibility (Python-native APIs that make customization straightforward).
From Pendulums to Humanoids: Example-Driven Design
Newton ships with an extensive library of examples that demonstrate its breadth. Basic examples cover pendulums, URDF robot loading, joint systems, conveyor belts, and height fields. But the robot examples are where things get genuinely exciting.
Users can simulate the Unitree G1 and H1 humanoid robots, the ANYmal C and D quadrupeds, the UR10 industrial arm, the Panda hydrodynamic arm, and even the Allegro Hand — all with a handful of Python commands. A policy learning example shows how to train RL controllers on these simulated robots, bringing the full pipeline from simulation to learned behavior within reach of any Python developer.
The cable examples — covering flexible cables, chains, and deformable objects — point to applications beyond traditional robotics: surgical simulation, cable routing in industrial design, and real-time physics for gaming and visual effects.
A Cross-Platform Story
Newton runs on Linux (x86-64 and ARM64), Windows, and macOS. On macOS, it falls back to CPU computation, but on any NVIDIA GPU-equipped Linux or Windows machine, the full GPU acceleration is available. Crucially, no local CUDA Toolkit installation is required — Newton ships as a pip-installable Python package that bundles everything needed.
System requirements are straightforward: NVIDIA GPU (Maxwell generation or newer), driver version 545 or newer (CUDA 12), and Python 3.10 or higher. Getting started is as simple as pip install newton[examples] followed by python -m newton.examples basic_pendulum.
Why Differentiability Matters
One of Newton is most distinctive features is its differentiable physics layer. Unlike traditional physics engines that merely simulate, differentiable physics engines compute gradients — they can tell you how the output of a simulation would change if any input parameter were tweaked. This is foundational for reinforcement learning, system identification, and model-predictive control.
In practice, this means researchers can define a physics simulation, define a loss function over the simulation outcomes, and use gradient-based optimization to automatically find physical parameters that match real-world robot behavior — a technique known as system identification. Or they can train neural network controllers that learn from physics gradients rather than purely from trial and error.
The Industry Coalition Behind It
Newton’s pedigree is remarkable: Disney Research (known for its visual effects and robotics research), Google DeepMind (behind MuJoCo and a long history of robotics simulation), and NVIDIA (whose Warp framework and CUDA ecosystem power the underlying compute). All three have contributed not just name recognition but core technology: MuJoCo Warp is Newton’s primary backend, Warp provides GPU primitives, and NVIDIA’s CUDA optimization ensures the engine scales across GPU generations.
The Linux Foundation hosting means Newton is unlikely to disappear or go commercial in unexpected ways — an assurance that enterprise adopters and academic researchers both need when committing to a simulation platform.
Looking Forward
Robotics simulation has long been a bottleneck in the pipeline from research to product. Newton addresses that bottleneck directly: open, GPU-native, differentiable, and backed by industry leaders. Whether it becomes the de facto standard for robot simulation or simply raises the bar for what open simulation can achieve, it represents a significant step toward closing the sim-to-real gap that has plagued robotics for decades.
For simulation researchers, robotics engineers, and anyone building systems that need physically accurate GPU-accelerated dynamics, Newton is worth watching — and worth trying. Install it, run the examples, and see what a modern physics engine feels like when it is designed from the ground up for the age of AI.