Nvidia tools accelerate humanoid robots and physical AI
Nvidia is rolling out a new suite of tools aimed at accelerating humanoid robotics and “physical AI,” unveiled this week at the Conference on Robot Learning (CoRL) in Seoul, South Korea.
What’s new: a three-part stack for humanoids
The company introduced three key upgrades: the open-source Newton Physics Engine, a new version of its Isaac GR00T reasoning model (GR00T N1.6), and expanded Cosmos World foundation models to generate synthetic training data. Nvidia says the stack is designed to boost adoption of humanoid robots, create more consistent testing, and move skills learned in simulation into the real world faster.
Newton Physics Engine (open-source)
At the center is Newton, a physics engine co-developed with Google DeepMind and Disney Research. It aims to give researchers more realistic, contact-rich simulation so robots can practice complex behaviors—such as walking over gravel or snow and manipulating small or deformable objects—before attempting them outside the lab.
Isaac GR00T N1.6
GR00T N1.6, the latest iteration of Nvidia’s robot reasoning model, is intended to bring more human-like problem solving to embodied systems. Available on Hugging Face, it integrates with Nvidia’s Cosmos Reason vision-language model, which the company says can turn ambiguous instructions into step-by-step action plans by drawing on prior knowledge, common sense, and basic physics, helping robots generalize across tasks.
Cosmos World foundation models
To support large-scale training, Nvidia also updated its open Cosmos foundation models. Developers can now generate diverse synthetic datasets from text, image, and video prompts to train and validate physical AI systems at scale.
Why it matters: sim-to-real, standardized testing
Nvidia frames the combination of GR00T, Newton, and its Omniverse simulation platform as providing the cognition, embodiment, and training environment needed to take humanoid robots from research prototypes to everyday use. The goal is to speed sim-to-real transfer and create more consistent, repeatable benchmarks for humanoid capabilities.
Who’s trying it first
Early adopters include robotics companies Lightwheel, Neura Robotics, and LG Electronics, which are using GR00T N1.6 for humanoid development. Research groups such as ETH Zurich’s Robotic Systems Lab, the Technical University of Munich, and Peking University are also among the first to trial the tools.
The big picture
Together, Newton’s contact-rich simulation, GR00T’s task reasoning, and Omniverse’s training environment aim to standardize development workflows and accelerate the path from simulation to real-world deployment for humanoid robots and broader physical AI systems.
