Newton: physics simulation engine built upon NVIDIA Warp (github.com)

Newton is a new GPU-accelerated physics simulation engine, now in active beta, built on NVIDIA Warp and using MuJoCo Warp as its primary backend. Targeted at roboticists and simulation researchers, Newton extends Warp’s deprecated warp.sim module to provide a modern, GPU-first simulation stack that emphasizes differentiability, OpenUSD export, and user-defined extensibility. The project is experimental (unstable API and likely breaking changes), but already ships a wide array of examples—articulations, URDFs, IK benchmarks, cloth, MPM granular materials, soft bodies and drone diffsims—that can be launched via the uv runner with configurable viewers (OpenGL, USD, ReRun, or null), device selectors (cpu, cuda:0, etc.), and other runtime flags. For the AI/ML community, Newton’s significance is twofold: it makes large-scale, GPU-native physical simulation and differentiable physics more accessible for gradient-based control, sim-to-real research, and learned dynamics, and it integrates OpenUSD output for easier visualization and pipeline interoperability. Practically, researchers get a fast, extensible backend that supports Torch-enabled examples (e.g., torch-cu12), USD export for downstream tooling, and hooks for custom operators—speeding iteration for model-based RL, system ID, and differentiable optimization workflows. Expect breaking API updates while the team refines design and governance under the Linux Foundation code-of-conduct.