Introduction

Simulating physics is central to robotics: before a robot ever moves in the real world, much of its learning, testing, and control happens in a virtual environment. But traditional simulators often struggle to match real-world physical complexity, especially where contact, friction, deformable materials, and unpredictable surfaces are involved. That discrepancy is known as the sim-to-real gap, and it’s one of the biggest hurdles in robotics and embodied AI.

On September 29th, the Linux Foundation announced that it is contributing Newton, a next-generation, GPU-accelerated physics engine, as a fully open, community-governed project. This move aims to accelerate robotics research, reduce barriers to entry, and ensure long-term sustainability under neutral governance.

In this article, we’ll unpack what Newton is, how its architecture stands out, the role the Linux Foundation will play, early use cases and challenges, and what this could mean for the future of robotics and simulation.

What Is Newton?

Newton is a physics simulation engine designed specifically for roboticists and simulation researchers who want high fidelity, performance, and extensibility. It was conceived through collaboration among Disney Research, Google DeepMind, and NVIDIA. The recent contribution to the Linux Foundation transforms Newton into an open governance project, inviting broader community collaboration.

• GPU-accelerated simulation: Newton leverages NVIDIA Warp as its compute backbone, enabling physics computations on GPUs for much higher throughput than traditional CPU-based simulators.

• Differentiable physics: Newton allows gradients to be propagated through simulation steps, making it possible to integrate physics into learning pipelines (e.g. backpropagation through control parameters).

• Extensible and multi-solver architecture: Users or researchers can plug in custom solvers, mix models (rigid bodies, soft bodies, cloth), and tailor functionality for domain-specific needs.

• Interoperability via OpenUSD: Newton builds on OpenUSD (Universal Scene Description) to allow flexible data modeling of robots and environments, and easier integration with asset pipelines.

• Compatibility with MuJoCo-Warp: As part of the Newton project, the MuJoCo backbone is adapted (MuJoCo-Warp) for high-performance simulation within Newton’s framework.