Newton
Open GPU physics engine for robot learning, with Kamino for closed-loop linkages
License
- code: Apache 2.0
- documentation: CC BY 4.0
Code: Apache-2.0, per LICENSE.md in newton-physics/newton and the PyPI package metadata. Documentation: CC BY 4.0, as the README states. Robot models and scenes are downloaded separately from the newton-assets repository and keep their own licences, mostly Apache-2.0, BSD or MIT; the Disney Research DR Legs and test-mechanism assets used by two Kamino examples may be used only for research or software development. Checked 2026-11-08 against LICENSE.md, the README, PyPI and the asset licence files.
Overview
Newton is an open-source physics engine for robotics and simulation research that runs on the GPU. It was started by Disney Research, Google DeepMind and NVIDIA and is now a Linux Foundation project, built on NVIDIA Warp and using MuJoCo Warp as its main rigid-body backend. One Python API gives access to several solvers, for rigid robots, cloth, cables, soft bodies and granular material, and robots load from URDF, MJCF or OpenUSD files. Isaac Lab can use Newton as its physics backend for reinforcement learning. Its Kamino solver, from Disney Research and presented at CoRL 2026, simulates robots with closed kinematic loops, such as parallel legs and four-bar linkages, directly on the GPU instead of approximating them as a tree. Newton is Apache-2.0 and installs from PyPI.
Newton is a GPU physics engine for roboticists and simulation researchers. It was started by Disney Research, Google DeepMind and NVIDIA, and in 2025 it was contributed to the Linux Foundation, which now hosts it under open governance. Newton is written in Python on top of NVIDIA Warp, which compiles Python kernels for the GPU, and it grew out of Warp's older simulation module. Its main rigid-body backend is MuJoCo Warp, the GPU version of MuJoCo that Google DeepMind and NVIDIA build together.
The idea is one model and one simulation loop with interchangeable solvers. Robots are built in Python or imported from URDF, MJCF or OpenUSD files, and the same model can be stepped by the MuJoCo solver, by Featherstone, XPBD or semi-implicit solvers for articulated bodies, by VBD for cloth, cables and soft bodies, by an implicit MPM solver for sand, snow and other granular material, by Style3D for garments, or by Kamino. Several solvers can be coupled in one scene, some support differentiable simulation for gradient-based optimisation, and the engine includes collision detection, sensors such as cameras and IMUs, inverse kinematics and viewers that run on the desktop or in the browser. It installs with pip and ships with more than a hundred runnable examples, from pendulums and walking quadrupeds to cloth and granular flows. Isaac Lab can run its environments on Newton as well as on PhysX, so reinforcement learning tasks written for Isaac Lab can use Newton's solvers; the community projects using it include RoboSimStudio.
Kamino: closed kinematic loops on the GPU
Most GPU simulators used for reinforcement learning assume a robot is a kinematic tree. Many real mechanisms are not: parallel legs, four-bar linkages in knees and grippers, and parallel manipulators form closed loops, and they are usually simulated by cutting each loop into a tree and adding a closure constraint or a mimic joint, which needs tuning and can change how the robot behaves. Kamino, developed at Disney Research in Zurich with NVIDIA and described in a CoRL 2026 paper, is a Newton solver that simulates such mechanisms as they are. Every body keeps its own pose, joints, joint limits and frictional contacts are solved together as hard constraints with a proximal ADMM method, and a forward kinematics solver puts closed-loop robots into a consistent pose when an episode resets. Parallel worlds on the GPU can also hold different robots, so one policy can be trained across several body shapes at once.
In the paper, Kamino trained a walking policy for DR Legs, a biped with six nested kinematic loops, using 4,096 parallel environments on one GPU, and on elementary tests with known analytical answers, such as a sliding box, a bouncing ball and a four-bar linkage, it matched them with little tuning, where the MuJoCo Warp and XPBD solvers needed tuning or showed penetration, wobbling or premature sliding. The authors also trained policies for Disney's BDX and Olaf robots in Kamino and report that they transferred to the physical robots (sim-to-real). Kamino is selected like any other Newton solver, the DR Legs model and a four-bar mechanism are among its examples, and the Kamino guide explains its three dynamics backends.
Why it matters
Newton is a physics layer that Disney Research, Google DeepMind and NVIDIA build together in the open, under neutral governance, rather than each keeping a separate engine. For a researcher that means one Apache-2.0 Python package where MuJoCo's GPU solver, cloth, cable and granular solvers and newer research solvers sit behind one interface, and Isaac Lab can train on it without a full Isaac Sim install. It sits next to Genesis, another GPU engine with many solvers, and mjlab, which puts the Isaac Lab API on MuJoCo Warp alone. Kamino adds native support for closed linkages: robots with them, which are common in legged robots and grippers, can be trained with reinforcement learning on the GPU without first being rebuilt as a tree. It is the more accurate and the slower option, and still experimental, so it suits mechanisms that need it rather than every robot.
Problem it solves
GPU robot simulators have been tied to one company's engine and one physics method, and most of them cannot simulate closed kinematic loops without approximations and tuning. Newton puts several solvers behind one open Python interface on the GPU, and its Kamino solver simulates parallel legs, linkages and other closed-loop mechanisms natively, at the scale needed for reinforcement learning.
Key takeaways
- GPU physics engine in Python, built on NVIDIA Warp, with MuJoCo Warp as its main rigid-body backend
- Started by Disney Research, Google DeepMind and NVIDIA, now hosted by the Linux Foundation
- One model and loop for many solvers: MuJoCo, Featherstone, XPBD, VBD, implicit MPM, Style3D and Kamino
- Imports robots from URDF, MJCF and OpenUSD; some solvers support differentiable simulation
- Isaac Lab can use Newton as its physics backend for reinforcement learning
- Kamino (CoRL 2026) simulates closed kinematic loops and hard contacts natively on the GPU
- In the paper, Kamino trained a walking policy for a biped with six nested loops in 4,096 parallel environments
- Apache-2.0, installs from PyPI, with more than a hundred runnable examples
Who it is for
- Robot learning researchers training policies in massively parallel GPU simulation
- Designers of legged robots, grippers and other mechanisms with parallel linkages or four-bar joints
- Isaac Lab and MuJoCo users who want a GPU backend with more solver choices
- Simulation researchers who want to add their own solver to a shared, open engine
- Teams simulating cloth, cables, soft bodies or granular material together with robots
Limitations
- Kamino is marked experimental and in beta: the documentation discourages depending on it and says its API and behaviour may change without notice; VBD is experimental too
- Kamino is slower than the MuJoCo Warp and XPBD solvers: in the paper it ran about 25 times fewer samples per second than MuJoCo Warp on the BDX robot, so it pays off only for mechanisms with closed loops
- Of the paper's robots only DR Legs is released; the BDX, Olaf and Iron Man models are proprietary, and the sim-to-real results on BDX and Olaf are the authors' own and, by their account, modest
- GPU acceleration needs an NVIDIA GPU of the Turing generation or newer with a recent driver; macOS runs on the CPU only
- The GPU solvers in the Kamino paper use 32-bit floating point, while MuJoCo on the CPU uses 64-bit
- A young, fast-moving project: version 1.0 came out in 2026, minor releases follow about monthly and the change log regularly deprecates parts of the API
- Robot models are downloaded separately and keep their own licences; the Disney Research assets used by two Kamino examples, DR Legs and a test mechanism, may be used only for research or software development