When AI began to reason, plan, and act inside digital systems, it marked the beginning of our journey toward agentic AI. These systems could make their own choices and complete tasks without constant human direction. But as businesses focus on releasing swarms of AI agents, NVIDIA is asking what comes next, and it's something it calls Physical AI.
Physical AI is the idea of taking intelligence out of servers and into the physical world. It powers robots, vehicles, drones, and machines that can sense and respond to their surroundings. The shift from agentic to physical intelligence moves AI from thought to action, where agentic AI analyzes and physical AI acts.
This new focus brings together NVIDIA's most advanced technologies. Through its Omniverse platform, world foundation models in Cosmos, and the Isaac robotics ecosystem, the company is laying the groundwork for machines that learn safely in synthetic worlds before ever moving into real ones. It is an ambitious attempt to create intelligence that can both reason and interact with reality.
undefinedPhysical AI and robotics will bring about the next industrial revolution.” - Jensen Huang.
undefined NVIDIA Omniverse (@nvidiaomniverse) October 1, 2025
Join us at Physical AI and Robotics Day at #NVIDIAGTC Washington, D.C. for expert-led sessions and hands-on trainings.
Don't miss highlights including:
🦾 Accelerating the Physical AI… pic.twitter.com/Gq6iUamz0d
Defining physical AI
Physical AI is made up of intelligent systems that don't just process data or generate language, they also analyze and interpret it. This means they understand concepts such as motion, weight, light, and texture. They learn how objects move and how forces interact with each other. But training this kind of intelligence requires more than large datasets.
City of Tokyo released a 3D digital twin of the entire city in high resolution point cloud, free to download. It’s an inevitable trend that more and more cities, houses, and factories will be transported into simulations.
undefined Jim Fan (@DrJimFan) December 24, 2024
Robots will not be trained in isolation. They will be… pic.twitter.com/IU5RxpFOfa
Omniverse is NVIDIA's platform for building and connecting these virtual worlds. This is built around OpenUSD, the Universal Scene Description (USD) format that was initially developed by Pixar.
OpenUSD defines every part of a 3D environment, from lighting and materials to physical behavior. This technology makes it easier for engineers to design digital twins of entire factories, energy plants, or cities. Scientists can also work within the same environment in real-time. This shared infrastructure enables the simulation of the physical world at scale.
Running Omniverse on NVIDIA's DGX Cloud platform lets organizations create these simulations without maintaining their own massive computing clusters. Omniverse is more than a design tool. It is the engine room of Physical AI.
How synthetic data accelerates training
Robots must see thousands of examples to perform reliably, but gathering that information safely may take months. Synthetic data changes that equation altogether.
Omniverse helps developers simulate factories, warehouses, and city streets with perfect detail. It then generates the sensor data that a robot would collect, such as labeled images, depth maps, and LiDAR reflections. Synthetic datasets can be tailored to specific tasks and edge cases that are too rare or dangerous to capture in our physical world.
The Omniverse Replicator framework and a process called domain randomization ensure that AI models do not overfit to a single environment. They learn to handle variations in light, movement, and perspective, which leads to better real-world performance once deployed.
To make these synthetic worlds trustworthy, NVIDIA promotes SimReady assets. These are 3D models created using OpenUSD that include physics and semantic information alongside their visual details. Each asset is more than an image. It contains the rules that define how it behaves when acted upon by gravity, light, or mechanical force.
The move toward standardized assets is supported by the Alliance for OpenUSD, a collaboration between NVIDIA, Pixar, and other partners. The goal is to make 3D content interoperable across all platforms.
For physical AI, this means that a robot trained in one digital twin can easily adapt to another without having to start from scratch. Standardization is what turns individual experiments into a scalable ecosystem.
The role of Isaac Sim and Isaac Lab
Training robots requires both accurate physics and intelligent feedback. NVIDIA's Isaac Sim and Isaac Lab handle these two needs. Isaac Sim is a simulation environment built on Omniverse that enables developers to test robots in highly detailed virtual spaces. Isaac Lab is a learning framework that allows robots to practice through reinforcement and imitation learning.
With Isaac Sim, developers build entire environments such as warehouses or offices. They can add SimReady assets for objects and surfaces, providing robots with a learning environment. The platform also integrates with tools such as MobilityGen, which generates motion trajectories for mobile robots, and NuRec, which reconstructs 3D scenes from real-world data using only a smartphone.
Isaac Lab then takes these virtual setups and turns them into training grounds. Robots practice picking up objects, navigating obstacles, or coordinating with other robots. Each trial produces new data that strengthens the model. Over time, this creates policies that can be transferred to physical machines with minimal adjustment.
The wait is over. 🥁
undefined NVIDIA Robotics (@NVIDIARobotics) August 11, 2025
NVIDIA Isaac Sim 5.0 and Isaac Lab 2.2 are now generally available.
Dive in and start to simulate, train, and validate your AI-powered robots with these new key features:
⚡ Instant GPU access via NVIDIA Brev
🤖 3D #GaussianSplatting with @NVIDIAOmniverse… pic.twitter.com/OkHmAeUdVB
Skild AI and Serve Robotics
Two examples show what Physical AI looks like outside the lab. Skild AI has developed an omnibodied foundation model that serves as a universal brain for robots. Instead of building separate models for each machine, Skild's system can control a wide range of different robot types. It was trained in simulation using NVIDIA's Isaac Lab and Cosmos Transfer to expand the diversity of its data.
Serve Robotics also uses NVIDIA's Jetson Orin hardware and Isaac Sim software to train and operate its sidewalk delivery robots. This resulted in 100,000 deliveries with an impressive 99.8% success rate, and now plans to scale to 2,000 robots. Their work shows how simulation can move from theory to real-world performance.
Cosmos and the emergence of world models
While Isaac handles simulation and control, Cosmos handles imagination. It is NVIDIA's environment for building World Foundation Models, a new class of generative models that create and understand 3D environments. Cosmos can generate photorealistic worlds from text prompts, videos, or sensor data, extending the capabilities of generative AI beyond the screen.
Cosmos Predict and Cosmos Transfer help developers build or modify synthetic worlds. By tweaking data around lighting, material, and geometry, they can produce endless variations of the same environment. This provides AI systems with a broader range of experiences, enabling them to learn and adapt to various scenarios.
The integration between Cosmos and Omniverse creates a complete loop. Cosmos builds worlds, Omniverse simulates them, and Isaac teaches robots to operate inside them. This continuous cycle of generation, simulation, and learning defines how NVIDIA envisions scaling physical AI for real-world use.
With 3M+ downloads and counting, NVIDIA Cosmos is redefining physical AI.
undefined NVIDIA AI (@NVIDIAAI) September 29, 2025
Announced at #CORL25, new Cosmos updates are allowing developers to generate diverse data for accelerating training robot models at scale.
👏 Cosmos Predict 2.5 will combine three models into one… pic.twitter.com/DyeedTWxgu
Reinforcement and imitation learning
For AI to operate in the physical world, it must go beyond recognizing patterns. It has to develop behaviors through practice. Reinforcement learning allows this by rewarding successful actions and penalizing mistakes. Over time, the agent develops strategies that yield better results.
In NVIDIA's ecosystem, Isaac Lab provides the foundation for this process. Robots perform thousands of tasks virtually, learning to move efficiently, handle unexpected changes, and recover from errors. They can test failure scenarios without risk to hardware. This accelerates the learning curve and produces robust behavior.
Imitation learning complements this process. By observing human demonstrations, robots can mimic expert performance. Combining imitation and reinforcement learning gives machines both direction and freedom. It is how they move from simple automation toward adaptive intelligence that can handle new challenges.
However, trust remains a major challenge. As AI systems become more autonomous, their decision-making processes must be transparent. Developers and regulators need ways to validate the safety of robots that learn from simulation.
NVIDIA's open approach aims to address these concerns by encouraging collaboration through the Alliance for OpenUSD and expanding enterprise access through DGX Cloud. The company aims to make safety and interoperability the foundation of scale, rather than an afterthought.
Why this shift matters
Physical AI represents more than a new product category. It marks a change in how industries will operate. Digital twins enable organizations to test designs, optimize operations, and train intelligent systems before deployment. This reduces cost, improves safety, and accelerates innovation.
🧵What if you could build, simulate, and collaborate on any 3D world—digitally and physically—to power the future of industry and robotics?
undefined NVIDIA AI (@NVIDIAAI) August 6, 2025
Aaron Luk, Director of Product Management for Simulation Tech at NVIDIA, explains how #OpenUSD is making this possible. 🌟 pic.twitter.com/TnMu3qW8o8
Factories can model production lines virtually before installing machines. City planners can simulate infrastructure stress and energy use. Healthcare systems can test hospital workflows without disrupting patient care. Physical AI connects these fields through a shared foundation of simulation and intelligence.
For NVIDIA, this is an opportunity to provide the infrastructure for a new era of computing. Just as GPUs powered the rise of deep learning, Omniverse and its connected platforms may become the backbone for intelligent physical systems. The company's bet is that simulation will soon be as essential to AI as data has been to software.
From agentic to physical intelligence
Agentic AI taught systems to reason. Physical AI is teaching them to exist. NVIDIA's work shows a clear path from abstract intelligence to embodied intelligence that can operate safely and intelligently in the real world.
NVIDIA and America’s leading manufacturing and robotics companies are driving U.S. reindustrialization with physical AI. 🙌
undefined NVIDIA Omniverse (@nvidiaomniverse) October 29, 2025
Learn how @agilityrobotics, @amazon Robotics, @Figure_robot, and @SkildAI are building America’s collaborative robot workforce using NVIDIA’s… pic.twitter.com/8hYTKdqZ6z
The shift from data to experience changes what AI means for industry and society. Robots trained through millions of simulated scenarios can perform complex tasks with confidence. Digital twins of cities and factories can help solve problems before they occur. The goal is not simply to automate, but to create systems that understand the context within which they operate.
NVIDIA's investments in Omniverse, Cosmos, and Isaac reflect a belief that intelligence must have a physical presence. It must see, move, and adapt. The next decade may reveal whether machines can truly learn to live in the world we have built for them.
Unlock more exclusive Cybernews content on YouTube.
Your email address will not be published. Required fields are markedmarked