← Back
NVIDIA
NVIDIA releases Newton 1.0, GPU-accelerated physics engine for robotics with 252x MuJoCo speedup
· releaseplatformperformanceopen-source · developer.nvidia.com ↗

Newton 1.0 GA Released

NVIDIA has released Newton 1.0 GA, a production-ready, GPU-accelerated physics engine designed specifically for robotic simulation. Built on NVIDIA Warp and OpenUSD, Newton provides a unified framework for both rigid-body and deformable simulation, enabling faster training and more realistic modeling of robotic tasks.

Key Features

Versatile Solver Architecture: Newton ships with multiple complementary rigid-body solvers:

  • Kamino (developed by Disney Research) handles complex mechanisms like robotic hands and legged systems with closed-loop linkages
  • MuJoCo 3.5 (MJWarp) now offers dramatic performance improvements: 252x speedup for locomotion and 475x for manipulation tasks on NVIDIA RTX PRO 6000 Blackwell Series GPUs

Advanced Collision and Contact Modeling:

  • Signed distance field (SDF)-based collision detection directly from CAD meshes, eliminating mesh approximation needs for tight-tolerance tasks like connector insertion
  • Hydroelastic contacts with continuous pressure distribution for high-fidelity object interaction and better sim-to-real transfer
  • Flexible collision detection pipeline supporting both broadphase and narrowphase approaches

Rich Deformable Simulation: Newton handles multiple material types through specialized solvers:

  • Vertex Block Descent (VBD) solver for cables, cloth, and rubber parts
  • Implicit Material Point Method (iMPM) for particle simulation in rough terrain scenarios
  • Deformable solvers couple explicitly with MuJoCo Warp for complex manipulation and locomotion

Integration and Usability

Newton provides a stable, unified API supporting multiple robot description formats (MJCF, URDF, and OpenUSD), making it easier to connect existing robot assets. The framework integrates natively with NVIDIA Isaac Sim 6.0 and Isaac Lab 3.0, enabling faster workflows from robot description through training and evaluation for both reinforcement and imitation learning.

The modular design allows teams to mix and match collision detection, contact models, sensors, control, and solver backends while maintaining a consistent simulation stack, significantly reducing development complexity for industrial robotics applications.