A physics toolkit for robot-scale parallelism
Newton is not a replacement for a game engine with a scene editor and a deploy button. It is a Python simulation library for roboticists and simulation researchers who need to construct physical worlds, step them quickly, and inspect or differentiate the results. The project builds on NVIDIA Warp, incorporates MuJoCo Warp as its primary backend, and grew out of work by Disney Research, Google DeepMind, and NVIDIA. It now sits under Linux Foundation governance with an Apache-2.0 code license.
Its central abstraction is straightforward. A ModelBuilder creates bodies, joints, shapes, and materials from code or imported assets. Finalizing produces a model with state and control buffers. A collision pipeline generates contacts, and a chosen solver advances state. Newton can replicate one robot template into many isolated worlds so a GPU handles a batch in parallel, the pattern needed for reinforcement learning and controller evaluation.
The scope goes well beyond rigid robots. Included examples cover articulations, cloth, particles, material point method simulations, cables, soft bodies, multiphysics contact, inverse kinematics, sensors, and several solver families. URDF, MJCF, and USD importers reduce the cost of bringing in existing assets. Viewers can display through OpenGL, write USD, use RTX, or connect to Rerun and Viser, while a null viewer keeps headless runs clean.
Installation is simple, dependency choices are not
Installing the examples extra and running python -m newton.examples is an unusually friendly entrance for a research simulator. The launcher exposes many runnable scenes, and common flags select the viewer, device, frame count, and USD output. The base package is smaller, requiring only Warp, and the installation guide includes a complete sphere-and-ground example that uses XPBD without pulling in the full example stack.
From there, users need to choose deliberately. The sim extra adds MuJoCo and MuJoCo Warp. Importers bring mesh processing and USD packages. Other extras add remeshing, ONNX inference, notebooks, or CUDA-specific PyTorch builds. The examples extra includes simulation, import, ONNX, and visualization dependencies, so its convenient command is heavier than the core package. Use a virtual environment and pin the dependency set that produced a result.
Hardware is the larger constraint. GPU acceleration requires a supported NVIDIA GPU and driver 545 or newer for CUDA 12, with newer versions recommended. No local CUDA toolkit is needed because Warp supplies its runtime. Linux and Windows can use the GPU, but macOS is CPU-only. The compatibility page tests Ada Lovelace and Blackwell devices even though inherited Warp support reaches older architectures, so owners of older cards should measure their workload rather than assume equal coverage. ARM64 Linux users also face GLIBC and X11 requirements for importer and example extras.
Breadth creates useful choices and validation work
A major advantage is the ability to compare approaches inside one model framework. Newton includes XPBD, VBD, MuJoCo, Featherstone, semi-implicit, Kamino, implicit MPM, and Style3D solvers. That makes it attractive to researchers whose work crosses rigid contact, deformables, cables, and learned control. It also avoids making one solver's assumptions the architecture of the entire application.
Not every component carries the same stability. Newton explicitly labels experimental features as subject to API, behavior, default, and support changes without notice. Version 1.5.0 added experimental joint controllers and optional Kamino dynamics while also changing defaults, removing older APIs, and announcing future removals. The documented deprecation window is only one full minor cycle, and only the latest minor release line receives fixes. A lab should regard minor upgrades as migration events, run regression scenes, and archive lockfiles with results.
Asset import deserves particularly strict tests. Open issue 3691 shows MJCF scaling changing geometry while leaving some linear joint and constraint quantities unscaled. Issue 3689 says the importer ignored MuJoCo's settotalmass, changing masses, inertia, contact, and controller behavior. Pull requests were already addressing parts of this work, which reflects an active project, but imported files should be compared with native MuJoCo on quantities important to the experiment. Visual similarity is not enough.
Current rough edges are scientifically meaningful
Issue 3897 is a good example of why validation matters. Its Newton 1.5.0 reproduction says XPBD heightfield contacts ignored elevation and treated terrain as a flat plane at the minimum height. The same report describes an MJCF path ejecting a nearby body at extreme speed on nonuniform terrain. Anyone training locomotion on heightfields should either verify a fix in the exact installed version or exclude that path.
Other open work includes MJCF fidelity, solver contact behavior, camera replacement, USD authoring gaps, and controller correctness. GitHub showed 374 open issues and pull requests combined, with 273 actual open issues. This is a large queue, but recent activity tells a healthier story than the number alone. Bug reports about importers and geometry frequently sit beside linked corrective pull requests, and both issues and code were moving daily.
The repository was pushed on August 14, 2026. Version 1.5.0 shipped on August 11, following 1.4.0 in July and a sequence of monthly feature releases since 1.0 in April. Release notes are exceptionally useful, calling out upgrade attention, removals, new deprecations, dependency changes, and measured improvements. The main documentation adds compatibility, migration, solver, concept, installation, and visualization guides around the extensive examples.
Newton is a strong choice when GPU batch simulation and solver experimentation outweigh the cost of a moving research platform. Start with a representative robot and contact scene, compare it against a trusted simulator, then pin everything. Teams needing slower API change, exact MuJoCo semantics, or non-NVIDIA acceleration should choose a more established or better-aligned alternative.