## Introduction Newton is an open-source, GPU-accelerated physics simulation engine built on NVIDIA Warp, specifically designed for roboticists and simulation researchers. With 3,000+ GitHub stars and an Apache 2.0 license, Newton provides a comprehensive multi-physics platform that supports rigid bodies, soft bodies, cables, cloth, granular materials, and articulated robots. The project extends the deprecated `warp.sim` module and integrates MuJoCo Warp as its primary computational backend, delivering the performance needed for large-scale robotics simulation and training. As embodied AI and robotic manipulation research accelerate, the need for fast, differentiable physics simulators has never been greater. Newton fills this gap with a Python-first interface that runs entirely on GPU, enabling researchers to simulate thousands of environments in parallel for reinforcement learning and gradient-based optimization workflows. ## Architecture and Design Newton's architecture leverages NVIDIA Warp for GPU-compiled physics kernels with a clean Python API: | Component | Purpose | |-----------|--------| | Rigid Body Solver | High-performance contact dynamics for rigid object simulation | | Soft Body Solver | Deformable body simulation with various material models | | Cable & Cloth | Specialized solvers for rope-like and fabric materials | | MPM Engine | Material Point Method for granular materials and fluids | | Articulation System | Multi-joint robot simulation with forward/inverse dynamics | | IK Solvers | Inverse kinematics for robotic manipulation planning | | Sensor Suite | Camera, IMU, and contact sensor simulation | | Differentiable Backend | Gradient computation through the physics simulation | | USD Pipeline | OpenUSD import/export for asset interchange with 3D tools | The engine requires Python 3.10+, NVIDIA GPU (Maxwell architecture or newer), and NVIDIA driver 545+ with CUDA 12. It also supports CPU execution for development and testing, though GPU acceleration is recommended for any serious workload. ## Key Capabilities **Multi-Physics Simulation**: Newton supports a broad range of physics phenomena in a single framework. Rigid bodies, soft bodies, cables, cloth, and granular materials (via Material Point Method) can all coexist in the same scene, enabling realistic simulation of complex robotic environments. **GPU Acceleration**: The entire simulation pipeline runs on GPU using NVIDIA Warp compiled kernels. This enables parallel simulation of thousands of environments simultaneously — critical for reinforcement learning approaches that require massive environment throughput. **Differentiable Simulation**: Newton supports computing gradients through the physics simulation, enabling gradient-based optimization of robot controllers, trajectory planning, and sim-to-real transfer. This differentiable pipeline is one of the key features for research workflows. **Inverse Kinematics**: Built-in IK solvers support robotic manipulation planning directly within the simulation environment, eliminating the need for external IK libraries and ensuring consistency between planning and simulation. **Sensor Simulation**: Camera rendering, IMU data, and contact force sensing are integrated into the simulation loop. This allows researchers to develop perception-action pipelines entirely within Newton before deploying to real hardware. **OpenUSD Support**: Newton uses Universal Scene Description (OpenUSD) for asset interchange, enabling seamless integration with 3D content creation tools, NVIDIA Omniverse, and other USD-compatible platforms. **Extensive Examples**: The repository includes 40+ runnable examples organized by domain: basic physics, robotics, cables, cloth, inverse kinematics, MPM, sensors, and differentiable simulation. Each example serves as both documentation and a starting point for new projects. ## Getting Started Newton provides straightforward installation via pip: ```bash # Basic installation pip install newton-sim # With example dependencies pip install newton-sim[examples] # Run an example python -m newton.examples.rigid_contact --device cuda:0 ``` Examples support command-line options for device selection (CPU/GPU), output format (OpenGL, USD, ReRun), and simulation parameters. ## Limitations Newton requires NVIDIA hardware for GPU acceleration, limiting accessibility for developers without compatible GPUs. The CUDA 12 and driver 545+ requirements may conflict with other GPU workloads in some environments. As an extension of the deprecated `warp.sim`, the migration path for existing users requires adaptation. The Python interface, while developer-friendly, adds overhead compared to native C++/CUDA implementations for the most performance-critical applications. Documentation, while extensive for examples, could be more comprehensive for the API reference and advanced customization. The project is actively evolving, meaning API stability is not yet guaranteed across versions. ## Who Should Use This Newton is built for robotics researchers who need fast, differentiable physics simulation for reinforcement learning, trajectory optimization, or sim-to-real transfer. Teams working on robotic manipulation benefit from the integrated IK solvers and contact dynamics. Researchers exploring multi-physics phenomena — such as robots interacting with deformable objects, cables, or granular materials — gain access to a unified simulation framework. NVIDIA ecosystem users already working with Warp, Omniverse, or Isaac Sim will find Newton a natural extension. Graduate students and research labs studying embodied AI can leverage the 40+ examples as an educational foundation and research platform.