Newton 1.0 GA: Production-Ready Physics Simulation for Robotics
NVIDIA announced Newton 1.0 GA at GTC 2026, a GPU-accelerated, open-source physics engine purpose-built for robotics simulation. As an extensible framework built on NVIDIA Warp and OpenUSD, Newton provides a unified architecture for handling complex dynamics including contact forces, deformable objects, and precise manipulation tasks.
Key Capabilities and Architecture
Newton is designed as a modular framework that unifies multiple solvers and simulation components:
- Stable API providing a consistent interface for modeling, solving, controlling, and sensing
- Multiple rigid-body solvers including MuJoCo 3.5 (MJWarp) and Kamino, enabling complementary capabilities for different use cases
- Rich deformable simulation via Vertex Block Descent (VBD) solver for cables, cloth, and volumetric objects, plus Implicit Material Point Method (iMPM) for particle simulation
- Advanced collision detection with signed distance field (SDF)-based collision for tight-tolerance tasks and hydroelastic contacts for realistic object interaction and tactile sensing
- Format flexibility supporting MJCF, URDF, and OpenUSD, making it easier to integrate existing robot assets and workflows
Performance Gains
The release delivers significant performance improvements on modern hardware:
- 252x speedup for locomotion tasks using MuJoCo Warp (MJX) on NVIDIA RTX PRO 6000 Blackwell Series
- 475x speedup for manipulation tasks with the same GPU architecture
- GPU-scale throughput enabling thousands of parallel training environments for reinforcement learning
Integration and Workflow
Newton integrates natively with NVIDIA Isaac Sim 6.0 and Isaac Lab 3.0, providing a faster pathway from robot description to trained policies. The common OpenUSD data layer enables seamless workflows across simulation, learning, and evaluation pipelines for both reinforcement and imitation learning.
Developer Action Items
Teams using robotics simulation should explore Newton as a foundation for dexterous manipulation and locomotion tasks. The framework's modular design allows mixing and matching solvers, collision detection methods, and custom components while maintaining a consistent simulation stack.