A new open-source physics simulation engine called Newton is making waves in the robotics community, offering GPU-accelerated simulations built on Nvidia Warp with differentiability and OpenUSD support baked in from day one.
The project, initiated by Disney Research, Google DeepMind, and Nvidia, extends Nvidia's deprecated warp.sim module and integrates MuJoCo Warp as its primary backend. Newton is now a Linux Foundation project maintained by the community under the Apache-2.0 license.
Built for Rapid Iteration
Newton emphasizes GPU-based computation, which dramatically accelerates physics simulations compared to traditional CPU-based approaches. The engine targets roboticists and simulation researchers who need to iterate quickly on designs, train reinforcement learning models, or run thousands of parallel simulations.
Key features include:
GPU acceleration via Nvidia Warp (CUDA 12, no local CUDA Toolkit required)
Differentiability for gradient-based optimization and deep learning integration
OpenUSD support for industry-standard scene interchange
User-defined extensibility for custom physics behaviors
MuJoCo Warp backend for proven physics algorithms
Comprehensive Examples
Newton ships with an extensive example library covering real-world use cases:
Robot simulations: CartPole, Unitree G1, Unitree H1, ANYmal C/D, UR10, Franka Panda, Allegro Hand
Soft body physics: Cloth simulation, cable dynamics, deformable objects
Inverse kinematics: End-effector control, cube stacking, custom IK solvers
MPM (Material Point Method): Granular materials, multi-material interactions
Differentiable simulation: Gradient-based optimization for control and design
Getting started is straightforward:
pip install "newton[examples]"
python -m newton.examples basic_pendulumWhy It Matters
Physics simulation is critical for robotics development. Training robot control policies in the real world is slow, expensive, and potentially dangerous. Simulation lets researchers test millions of scenarios safely and cheaply—but only if the physics are accurate and the simulations run fast enough.
GPU acceleration addresses the speed problem. A single high-end GPU can simulate thousands of parallel environments, accelerating reinforcement learning that might take weeks on CPUs to hours on GPUs. Newton builds on this with differentiability—the ability to compute gradients through the simulation itself.
Differentiable simulation enables gradient-based optimization of robot designs and control policies directly through physics, rather than relying solely on black-box reinforcement learning. This can dramatically reduce the sample complexity of learning tasks.
Requirements and Platform Support
Newton requires:
Python 3.10+
Linux (x86-64, aarch64), Windows (x86-64), or macOS (CPU only)
Nvidia GPU (Maxwell or newer) with driver 545+ for GPU acceleration
The project is community-maintained under the Linux Foundation, with contributions from major players in robotics and simulation. Documentation and detailed examples are available at the GitHub repository.
For robotics researchers tired of slow simulations or looking for differentiable physics, Newton offers a modern, GPU-first alternative backed by industry heavyweights and an active open-source community.
