← Back
NVIDIA
NVIDIA releases Newton 1.0 GA, GPU-accelerated physics engine for robotic simulation
· releaseplatformfeatureopen-sourceperformance · developer.nvidia.com ↗

Overview

NVIDIA has announced Newton 1.0 GA, a production-ready physics simulation engine optimized for robotic applications. Announced at NVIDIA GTC 2026, Newton is an open-source, GPU-accelerated simulator built on NVIDIA Warp and OpenUSD that enables realistic modeling of contact forces, complex dynamics, and deformable objects—critical for training robots on manipulation and locomotion tasks.

Key Features

Modular Architecture & Solver Options Newton provides a unified simulation framework with multiple interchangeable solvers:

  • MuJoCo Warp (3.5): Google DeepMind's trusted MuJoCo engine extended to GPU scale, achieving 252x speedups for locomotion and 475x for manipulation tasks on NVIDIA RTX PRO 6000 Blackwell GPUs
  • Kamino: Disney Research's solver for complex mechanisms (robotic hands, legged systems with closed-loop linkages) that handles passive actuation without simulatability constraints

Advanced Contact Modeling

  • Hydroelastic contacts: Continuous pressure distribution across finite-area contact patches instead of discrete points, providing higher-fidelity tactile sensing and manipulation policies
  • SDF-based collision detection: Captures complex CAD geometries directly without mesh approximation, enabling tight-tolerance tasks like connector insertion

Deformable Simulation Supports multiple material types through specialized solvers:

  • Linear deformables (cables) and thin deformables (cloth) via Vertex Block Descent (VBD)
  • Volumetric deformables (rubber parts) and granular materials via Implicit Material Point Method (iMPM)

Cross-Format Compatibility Unified runtime data model supporting MJCF, URDF, and OpenUSD formats, making it easier to import and work with existing robot assets and workflows.

Integration & Developer Access

Newton integrates natively with NVIDIA Isaac Sim 6.0 and Isaac Lab 3.0, enabling streamlined workflows from robot description to trained policies. A stable API provides a consistent interface for modeling, solving, control, and sensing across robot learning projects. The engine is available as open-source software built on the Warp framework.