NVIDIA Newton 1.0: Open-Source Physics Engine for Robotics Sim-to-Real

The biggest problem in robotics isn’t the robots. It’s the simulation.

You can train a robot arm for thousands of hours in a virtual environment and watch it fail the moment it touches a real cable, a real connector, or a real surface with real friction. That gap between simulation and reality — the sim-to-real gap — has been the quiet bottleneck slowing industrial robotics for a decade.

 

Newton 1.0, released at NVIDIA GTC 2026, is a direct attempt to close it.

What Newton Actually Is

Newton is an open-source, GPU-accelerated physics engine built on NVIDIA Warp and OpenUSD, developed jointly by NVIDIA, Google DeepMind, and Disney Research, and managed by the Linux Foundation. It’s free to use, modify, and extend.

 

The core job: simulate physics accurately enough that a robot policy trained in Newton works when it hits the real world.

It ships with multiple rigid-body solvers, including MuJoCo Warp and Kamino, a Vertex Block Descent solver for deformable simulation of cables, cloth, and volumetric materials, a signed-distance-field collision library, and hydroelastic contact modeling.

Those aren’t buzzwords. Each one solves a specific failure mode that has broken robot training for years.

The Four Technical Problems It Solves

1. Speed without sacrificing accuracy

Most physics engines make you choose. Fast simulation means shortcuts. Shortcuts mean the robot learns behaviors that don’t transfer to reality.

Newton clocks 475x faster than Google DeepMind’s MJX for manipulation tasks on RTX PRO 6000 Blackwell GPUs. That speed matters because more simulation time means more training trajectories, which means better policies — without sacrificing the physical accuracy that makes those policies transferable.

2. Closed-loop mechanisms

Disney Research’s Kamino solver, integrated into Newton, handles closed-loop mechanisms — robotic hands, legged systems with parallel linkages — that most physics engines simply cannot simulate reliably.

3. CAD-level collision detection

The collision detection system introduces signed distance field collision that ingests CAD meshes directly, eliminating the convex hull approximations that lose geometric detail on tight-tolerance parts. For connector insertion or board placement, that geometric fidelity is the difference between a policy that works and one that doesn’t.

4. Distributed contact pressure

Hydroelastic contact modeling generates distributed pressure across contact patches rather than point approximations — that distinction matters for connector insertion and tactile manipulation where real-world friction behavior determines success or failure.

Who Is Already Using It

This isn’t a research demo. Two production deployments are already running.

🔹Skild AI is using Newton with Isaac Lab to train reinforcement learning policies for GPU rack assembly — connector insertion, board placement, and fastening with tight tolerances.

 

🔹Samsung is using it for cable manipulation in refrigerator assembly lines. Cables are one of the hardest objects to simulate — they deform, self-collide, and change shape under force. Newton’s VBD deformable solver handles them well enough that Samsung is generating synthetic training data directly from the simulation.

🔹Lightwheel is codeveloping and calibrating the Newton physics engine to enable Samsung’s assembly robots to master intricate cable handling in simulation, delivering higher precision and faster assembly lines.

The Collaboration Behind It

Toyota Research Institute joined to advance solver development and contact modeling, bringing their work on the Drake physics engine into the project.

That combination matters for robotics simulation:

🔹Google DeepMind brings MuJoCo — the physics engine already trusted across robotics research.

🔹Disney Research brings Kamino and real-world character robotics experience.

🔹NVIDIA brings the GPU infrastructure and the Isaac Lab ecosystem that over 2 million robots worldwide already run on.

What NVIDIA Newton 1.0 Means for Builders

Newton plugs into Isaac Lab 3.0 and Isaac Sim 6.0 as a swappable physics backend. Teams author environments once, validate across different physics engines, then deploy.

That’s the practical unlock. You don’t rewrite your training pipeline. You swap the backend, run it faster and more accurately, and get policies that transfer better.

The robotics simulation market is approaching $28 billion. The companies building fluency in simulation-to-real pipelines now — understanding how physics engines, RL training loops, and real hardware connect — are positioning themselves at the infrastructure layer of physical AI.

Newton is free. The GitHub repo is live. The window to build on it before it becomes standard is open right now.

Get started: newton-physics/newton on GitHub · Isaac Lab 3.0 on GitHub

What’s Next?

Get more breakdowns like this in your inbox. Subscribe to The AI Entrepreneurs newsletter for weekly bite‑sized tutorials, tools, and playbooks to build smarter, faster, and with less guesswork. Join 70K+ founders and creators at AI Entrepreneurs — STANDOUT DIGITAL.