Production-Ready Physics Engine for Robotics
NVIDIA announced Newton 1.0 GA at GTC 2026, a fully production-ready GPU-accelerated physics simulator purpose-built for robotic manipulation and locomotion. The engine addresses a critical gap in robotics development by combining simulation speed with physical accuracy—enabling realistic modeling of contact forces, deformable objects, and complex mechanical systems without sacrificing performance.
Architecture and Modularity
Newton is built as a modular framework layered on NVIDIA Warp and OpenUSD, providing a unified API across multiple specialized solvers. Rather than forcing a single scene format, it supports MJCF, URDF, and OpenUSD—making it compatible with existing robot assets and workflows. Developers can mix and match collision detection, contact models, sensors, and solver backends while maintaining a consistent simulation stack.
Performance and Solver Capabilities
The release includes multiple rigid-body solvers with complementary strengths:
- MuJoCo Warp (MJWarp) delivers 252x speedup for locomotion and 475x for manipulation tasks on NVIDIA RTX PRO 6000 Blackwell GPUs compared to prior implementations, enabling GPU-scale training of thousands of parallel environments
- Kamino (from Disney Research) excels at complex mechanisms like robotic hands and legged systems with closed-loop linkages and passive actuation
Advanced Contact and Deformable Simulation
Newton introduces production-grade contact modeling critical for real-world transfer:
- Signed distance field (SDF) collision captures complex CAD-exported geometries precisely, eliminating mesh approximation errors for tight-tolerance tasks like connector insertion
- Hydroelastic contact models use continuous pressure distributions across contact patches rather than discrete points, improving tactile sensing and manipulation policy training
- Deformable solvers handle cables, cloth, rubber parts (via Vertex Block Descent), and granular materials through Implicit Material Point Method (iMPM)
Integration and Workflow
Newton integrates natively with NVIDIA Isaac Sim 6.0 and Isaac Lab 3.0 (early access), providing seamless workflows from robot description through policy training to evaluation. The OpenUSD-based architecture enables faster iteration across reinforcement and imitation learning pipelines. A Warp-based tiled camera sensor supports high-throughput rendering with RGB, depth, normals, and instance segmentation channels.
Key Takeaways for Developers
- Stable, unified API for modeling, solving, controlling, and sensing across robotics simulations
- Dramatic performance gains enable practical training of complex manipulation and locomotion behaviors
- Advanced contact modeling (hydroelastic, SDF) improves sim-to-real transfer for dexterous tasks
- Modular design allows mixing solvers and components for custom robotics workflows