NVIDIA has released the open source Newton Physics Engine within NVIDIA Isaac Lab, alongside the open NVIDIA Isaac GR00T N1.6 reasoning vision language action model. The update expands NVIDIA’s robotics ecosystem with new AI infrastructure that enables faster development, standardized testing, and consistent training-to-inference workflows. Together, these tools give developers an open, accelerated platform to build and test robot skills that transfer safely from simulation to the real world.
“Humanoids are the next frontier of physical AI, requiring the ability to reason, adapt and act safely in an unpredictable world,” said Rev Lebaredian, vice president of Omniverse and simulation technology at NVIDIA. “With these latest updates, developers now have the three computers to bring robots from research into everyday life – with Isaac GR00T serving as robot’s brains, Newton simulating their body and NVIDIA Omniverse as their training ground.”
Newton Opens New Standard for Physical Simulation in Robotics
Robots can learn tasks faster and with less risk in simulation, but humanoid robots – with their complex balance and motion – require accurate physics models. More than 250,000 robotics developers worldwide need simulations to ensure that skills trained virtually can be performed safely in real world.
NVIDIA has released the beta version of Newton, an open-source, GPU-accelerated physics engine managed by the Linux Foundation. Built on NVIDIA Warp and OpenUSD, and developed in collaboration with Google DeepMind and Disney Research, Newton is available to the global robotics community.
Newton’s flexible architecture supports multiple physics solvers, allowing developers to simulate robotic actions such as walking on uneven surfaces or handling cups and fruits.
Research institutions including ETH Zurich’s Robotic Systems Lab, Technical University of Munich, and Peking University have adopted Newton, along with Lightwheel and Style3D.
Cosmos Reason Improves Robot Reasoning for New Open Isaac GR00T N1.6 Model
To perform humanlike tasks in the physical world, humanoids should understand ambiguous instructions.
The new version of the open Isaac GR00T N1.6 robot foundation model will soon be available on Hugging Face. It will include NVIDIA Cosmos Reason, an open reasoning vision language model designed for physical AI. Cosmos Reason acts as the robot’s brain, converts vague instructions into step-by-step plans. It applies prior knowledge, basic reasoning, and physics to handle new situations and adapt to new tasks.
With more than one million downloads, Cosmos Reason ranks first on the Physical Reasoning Leaderboard on Hugging Face. It can also organize and label large and synthetic datasets for training AI models. Cosmos Reason 1 is available through the NVIDIA NIM microservice for AI model deployment.
Isaac GR00T N1.6 lets humanoids move and handle objects simultaneously, allowing more torso and arm freedom to complete tasks like opening heavier doors.
Developers can post-train Isaac GR00T N models using the open-source NVIDIA Physical AI Dataset on Hugging Face. Downloaded over 4.8 million times, the dataset includes thousands of synthetic and real-world trajectories.
Robot makers such as AeiROBOT, Franka Robotics, LG Electronics, Lightwheel, Mentee Robotics, Neura Robotics, Solomon, Techman Robot and UCR are evaluating Isaac GR00T N models for building general-purpose robots.
New Cosmos World Foundation Models for Physical AI Development
NVIDIA announced new updates to its open Cosmos WFMs, downloaded over 3 million times, that let developers generate diverse data for accelerating training physical AI models using text, image and video prompts.
- Cosmos Predict 2.5, coming soon, combines the power of three Cosmos WFMs into one model, reducing complexity, saving time and boosting efficiency. It supports longer video generation – capable of creating up to 30-second videos.
- Cosmos Transfer 2.5, coming soon, delivers faster, higher-quality results than previous models, while being 3.5x smaller. It can generate photorealistic data from 3D simulation scenes and spatial control inputs like depth, segmentation, edges and high-definition maps.
New Workflow for Teaching Robot Grasping
The developer preview of Isaac Lab 2.3, built on the NVIDIA Omniverse platform, introduces a dexterous grasping workflow that trains multi-fingered robots in virtual environments. The system uses an automated curriculum that begins with simple tasks and gradually increases complexity. It adjusts conditions such as gravity, friction, and object weight, helping robots develop adaptable skills for unpredictable situations.
Boston Dynamics’ Atlas robots learned grasping using this workflow to improve its manipulation capabilities.
Robot developers Agility Robotics, Boston Dynamics, Figure AI, Hexagon, Skild AI, Solomon and Techman Robot are adopting NVIDIA Isaac and Omniverse technologies.
Evaluating Learned Robot Skills in Simulation
Teaching a robot new skills, such as grasping objects or walking, is complex and costly when tested on physical machines. Simulation provides a safer, faster way to evaluate performance across environments and tasks.
However, many simulations remain too simple, limiting a robot’s ability to handle real-world conditions. To address this, NVIDIA and Lightwheel are jointly developing Isaac Lab – Arena, an open-source policy evaluation framework for large-scale testing and experimentation. The framework will allow developers to assess robotic behavior in diverse scenarios and will be available sooner.
New NVIDIA AI Infrastructure Powers Robotics Workloads Anywhere
To enable developers to take advantage of these technologies and software libraries, NVIDIA announced AI infrastructure designed for demanding workloads, including:
- NVIDIA GB200 NVL72, a rack-scale system integrating 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell GPUs, which is being adopted by cloud providers to accelerate AI training and inference, including reasoning and physical AI tasks.
- NVIDIA RTX PRO Servers, which offer a single architecture for every robot development workload across training, synthetic data generation, robot learning and simulation. RTX PRO Servers have been adopted by the RAI Institute.
- NVIDIA Jetson Thor, powered by a Blackwell GPU, allows robots to run multiple AI models simultaneously and perform on-device inference for physical AI applications, including humanoid robotics. Jetson Thor has been adopted by partners such as Figure AI, Galbot, Google DeepMind, Mentee Robotics, Meta, Skild AI and Unitree.
NVIDIA Advances Robotics Research
NVIDIA’s GPUs, simulation tools, and CUDA-accelerated libraries were cited in nearly half of the papers accepted at CoRL. The technologies are used by research institutions including Carnegie Mellon, University of Washington, ETH Zurich and National University of Singapore.
The conference also featured BEHAVIOR, a robotic learning project from the Stanford Vision and Learning Lab, and Taccel, a vision-based tactile robotics simulation platform developed by Peking University.
Source: NVIDIA
About NVIDIA
NVIDIA, founded in 1993 and headquartered in Santa Clara, CA, designs and manufactures graphics processing units, systems on chips, networking hardware, and AI intelligence software such as CUDA. Its products serve industries including gaming, data centers, autonomous vehicles, professional visualization, robotics, health care, and energy. The company introduced the GPU in 1999 and later expanded into accelerated computing and AI infrastructure. In gaming, its GPUs support high-performance rendering, while in AI and high-performance computing, its systems provide the infrastructure for training and deploying large-scale models. NVIDIA also develops tools for robotics and autonomous driving. For the fiscal quarter ending in July 2025, the company reported revenue of $46.7B and net income of $26.4B.