Nvidia drops open-source physics engine for robots

Newton is an open-source, GPU-accelerated physics engine co-developed by Nvidia, Google DeepMind, and Disney Research, managed by the Linux Foundation and available now in beta inside Nvidia Isaac Lab.

It is built to solve the primary bottleneck in humanoid robotics: transferring learned skills from simulation to reality without physical failure. Newton tackles complex dynamic contact, allowing researchers to accurately simulate humanoids walking across loose gravel and deep snow, or handling delicate objects like cups and fruit.

What you can do with it

Alongside Newton, Nvidia expanded its robot learning stack with the Isaac GR00T N1.6 foundation model, which will soon land on Hugging Face. Powered by the open Cosmos Reason vision-language model, GR00T N1.6 translates ambiguous instructions into step-by-step physical plans. Critically, it unlocks coordinated whole-body mobility, letting robots maneuver their torso and arms simultaneously to perform demanding tasks like opening heavy doors while in motion.

Manipulation training also gets an automated workflow. In the Isaac Lab 2.3 developer preview, an automated curriculum dynamically randomizes parameters like friction, gravity, and object mass to teach multi-fingered hands how to grasp unpredictable items. Boston Dynamics recently used this exact pipeline to upgrade manipulation skills on its Atlas humanoid.

For testing, Nvidia and Lightwheel are introducing Isaac Lab - Arena, an upcoming open-source framework designed for standardized policy benchmarking across simulated environments.

How it runs and specs

Newton is built on Nvidia Warp and OpenUSD, featuring a flexible solver architecture that integrates cleanly with platforms like Isaac Lab and MuJoCo Playground.

For post-training, developers can tap into the open Physical AI Dataset on Hugging Face, which has surpassed 4.8 million downloads. Nvidia is also rolling out Cosmos Predict 2.5 for generating up to 30-second multi-camera video simulations, alongside Cosmos Transfer 2.5, which generates photorealistic synthetic sensor data while being 3.5 times smaller than earlier releases.

On the hardware front, high-throughput simulation runs on GB200 NVL72 systems and RTX PRO servers, while real-time on-robot inference is handled by the new Blackwell-based Jetson Thor module.

Newton is available immediately in Isaac Lab, Cosmos Reason can be deployed via Nvidia NIM microservices, and GR00T N1.6 will arrive on Hugging Face shortly.

Why it matters

Robotics engineers and embodied AI labs should definitely test Newton, especially those struggling with contact stability, rough terrain locomotion, and sim-to-real transfer gaps. Having an engine backed by DeepMind, Disney, and the Linux Foundation provides a credible, community-governed alternative to fragmented proprietary setups.

We think this is a strategic play by Nvidia to own the physical AI toolchain. By open-sourcing the engine, foundation models, and evaluation suites, it pushes OpenUSD as the universal format for physical computing while ensuring that every step of the pipeline scales naturally on Nvidia infrastructure.