Skip to main content
Back

Disney, Google DeepMind, and NVIDIA's Newton Brings GPU-Powered Robot Simulation to Open Source

The Newton physics engine, a Linux Foundation project built on NVIDIA Warp, has crossed 5,000 GitHub stars as a community-maintained robotics simulator.

Repo, Full Access
  • Newton is a GPU-accelerated, differentiable physics engine for robotics, now past 5,000 GitHub stars.
  • Community-maintained Linux Foundation project initiated by Disney Research, Google DeepMind, and NVIDIA.
  • Built on NVIDIA Warp with MuJoCo Warp as the primary physics backend.
  • Handles rigid bodies, cloth, cables, granular materials, and soft bodies in one modular solver architecture.
  • Apache 2.0 licensed, installable via pip install "newton[examples]", OpenUSD-native for Isaac Lab compatibility.
  • Still alpha stage with rapidly evolving API and several hundred open issues on GitHub.

A community-maintained physics engine aimed squarely at robot learning has been quietly climbing the GitHub charts. The newton-physics/newton repository now sits above 5,000 stars, positioning itself as one of the more serious open-source alternatives for GPU-accelerated robotics simulation.

The project is unusual in its origins. The Linux Foundation welcomed Newton, an open source, GPU-accelerated, extensible physics engine that enables faster, more scalable simulations to lower the barrier to robotics research, co-developed by Disney Research, Google DeepMind, and NVIDIA and designed by and for robotics developers to tackle the challenges of building generalist robots. The three companies contributed the code, but the repository itself is now steered by the community under neutral governance.

What Newton actually is

Under the hood, Newton is a Python-first simulation stack layered on top of NVIDIA's Warp framework. The open-source Newton engine builds atop NVIDIA Warp as a developer framework for building and accelerating data generation and spatial computing in the Python programming language, and Warp in turn is accelerated when making use of NVIDIA CUDA-X. On top of that foundation, the project extends Nvidia's deprecated warp.sim module and integrates MuJoCo Warp as its primary backend.

If you have used MuJoCo before, that lineage matters. Newton is not throwing out a decade of contact dynamics work. It is wrapping a GPU-native execution model around it and adding features that were awkward or impossible in the classic CPU MuJoCo stack.