"The future of robotics isn’t just digital — it’s physical.”
That was the message from NVIDIA’s Deepu Talla during his keynote at Automate 2025. As one of the most influential voices in artificial intelligence (AI) and autonomy, he introduced us to a new era of robotics powered by physical AI. Talla outlined how simulation, generative AI, and large-scale computing converge to bring robots closer than ever to real-world intelligence and action.
Let's explore what physical AI really means and how it's leveraged in manufacturing environments today.
WHAT IS PHYSICAL AI VS. GENERATIVE AI?
Generative AI is a form of artificial intelligence designed to generate original outputs that mirror real-world inputs. Essentially, it creates new content (text, images, audio, or even code) by learning patterns from existing data.
Physical AI goes a step further for industrial applications. It’s used by machines that operate in the real, physical world. It enables autonomous systems like robots to perceive, reason, and act.
NVIDIA, a leader in this space, defines it as the use of vision-language-action models that can guide a robot’s decisions and movements.
Language → Vision → Vision-Language → Vision-Language-Action
WHAT IS FUELING PHYSICAL AI?
Physical AI is advancing quickly, but why now? It comes down to a powerful combination of market pressure and technological readiness.
Demand for flexible automation
Persistent labor shortages, aging infrastructure, and the urgent push to reshore manufacturing are all converging at once. For many companies, traditional automation is no longer sufficient, especially in high-mix, low-volume environments. What manufacturers need is autonomy. They need systems that can sense, adapt, and make decisions without constant reprogramming. That need is creating the perfect launchpad for physical AI.
Technology tipping points
Recent breakthroughs have pushed key enabling technologies past the threshold of usability. First, vision-language-action models are now capable of connecting perception, language, and decision-making for dynamic environments. At the same time, high-fidelity simulation platforms, such as NVIDIA’s Omniverse, have matured to the point where they can accurately reflect the real world for safer and quicker testing. There’s also synthetic data generation, which enables the training of these AI models at scale and eliminates the need for expensive real-world data collection.
Together, these tools are laying the foundation for the next wave of automation — intelligent systems that learn in virtual environments and adapt in the physical world.
From breakthrough to deployment
The leap from science fiction to functional robotics is happening now because of this convergence. Large-scale foundation models — specifically trained for robotics — are enabling robots to move from brittle, task-specific scripts to human-like behavior. With simulation tools that model complex environments and scalable data pipelines that train and refine behavior, robots are becoming more capable and reliable in real-world settings.
In other words, we’re no longer watching AI inch forward. Just as ChatGPT accelerated the adoption of generative AI across industries in a matter of months, Physical AI is on a similar step-function trajectory. This means that smarter, safer, and more autonomous machines can transition from the lab to the factory floor sooner than we think.
THE FOUR STEPS OF DEPLOYING PHYSICAL AI
To understand how physical AI becomes operational, Talla outlines a four-step cycle used by NVIDIA and its ecosystem partners.
Step 1: Create data
Data is everything, but real-world robotic data is scarce and expensive. That’s why companies are combining human demonstration, motion capture, and teleoperation with synthetic data generation using platforms like Cosmos and Omniverse.
Step 2: Train in simulation
Using tools like Isaac Lab, developers apply reinforcement learning and imitation learning to build robust AI models. In simulation, robots can fail safely and learn thousands of tasks at once.
Step 3: Test at scale
Digital twins enable full-scale testing before physical deployment. Want to crash-test a thousand robots in parallel? You can. Simulation is faster, safer, and cheaper — and now accurate enough to be predictive.
Step 4: Deploy in the real world
Once trained and tested, the AI is deployed to run on the robot itself. But it doesn’t stop there. Deployed robots continually feed back data to retrain and improve the model, forming a continuous learning loop.
INSIDE-OUT VS. OUTSIDE-IN AUTONOMY
NVIDIA also introduced a powerful concept for building smarter automation systems: inside-out and outside-in autonomy. To solve industrial automation, we need both of these.
Inside-out
Robots use embedded sensors to perceive their environment, acting like a robot’s eyes and ears.
Outside-in
Facility-level cameras and generative AI interpret what’s happening at scale and understand situations, like an air traffic control system.
EXAMPLES OF PHYSICAL AI FROM FACTORY TO HUMANOID
Physical AI is already here in both high-tech factories and emerging use cases in humanoid robotics. At Automate 2025, we saw how it’s helping companies bring intelligent autonomy to market today.
Industry collaborations
One standout example was Universal Robots’ UR15, which showcased real-time AI-driven motion generation and trajectory planning. Built using NVIDIA technology, the UR15 demonstrates how robots can move naturally, efficiently, and safely — even in complex environments.
Similarly, Vention introduced Machine Motion AI, aimed at helping small and midsize manufacturers implement automation without the cost and complexity that typically come with traditional systems. This signals a critical shift: physical AI is becoming more accessible for companies of all sizes.
NVIDIA’s reach extends into the control layer as well. KUKA and Standards Bots both unveiled how they integrate NVIDIA-powered AI directly into their robot controllers. This marks a move from AI as an “add-on” to AI as a core part of the robot’s brain to enable faster adaptation, smarter decision-making, and more flexible deployments.
Real-world examples
But physical AI isn’t limited to individual robots. It’s already playing a role in how entire factories are designed and operated. A prime example is Foxconn’s new GPU manufacturing plant in Guadalajara, Mexico, which was first built as a digital twin using NVIDIA Omniverse before installing a single piece of equipment on the floor. This approach allows companies to test layouts, workflows, and robotic processes entirely in simulation. This, in turn, significantly reduces development time and costs. As Deepu Talla notes in his keynote, “It’s free to break things in simulation.” It’s in those moments of failure that innovation can take off.
Humanoid robotics, once a distant vision, is also starting to take shape with help from Physical AI. Mercedes-Benz, in collaboration with Agility Robotics, is using simulation and AI to validate how humanoid robots might support future EV manufacturing lines. These systems are trained in virtual environments before entering production facilities, so manufacturers can test performance, safety, and coordination at scale.
From task-optimized arms to full-scale humanoids, the examples are clear. Physical AI is no longer theoretical. It's being tested, deployed, and scaled to help turn factories into autonomous ecosystems and push robotics into new frontiers.
WHAT’S NEXT? PHYSICAL AI IN 2026 AND BEYOND
As artificial intelligence matures, it becomes more accessible to manufacturers of all sizes. Even small and midsize facilities can start building intelligent automation systems using NVIDIA’s growing ecosystem of tools.
One of the most exciting developments is NVIDIA Isaac GR00T N1, the world’s first open humanoid robot foundation model. Designed with a dual-system architecture inspired by human cognition, it enables robots to both plan thoughtfully and act quickly. This paves the way for general-purpose autonomy across industries.
These models are also cross-embodied, meaning the same AI trained on a humanoid can be adapted to AMRs, robotic arms, or forklifts. That kind of versatility makes physical AI incredibly scalable, allowing manufacturers to reuse models across different form factors and applications.
Ease of use is another major shift. With natural language interfaces and low-code tools, deploying robotics is becoming less about writing scripts and more about describing outcomes. This dramatically lowers the barrier to entry for teams without deep AI or programming expertise.
As simulation platforms grow, so does the ability to test and coordinate multi-robot systems. From factory orchestration to logistics fleets, the future of physical AI is about intelligent, collaborative systems that learn, adapt, and evolve together. And that future is being built right now and may arrive faster than most expected.
SEE PHYSICAL AI IN ACTION AT AUTOMATE
Want to see what’s next for robotics and AI? Then you need to walk the Automate show floor, sit in expert-led conference sessions, and network with leaders paving the way for this technology.
At Automate, NVIDIA and other industry leaders are showing how to take the world’s most advanced automation technologies from concept to reality. Whether you're modernizing a single production line or designing systems from scratch, this is where you find the ideas, tools, and partners to make it possible.
Because the future isn’t coming, it’s already here. And it’s physical.
« View All Blogs
Privacy Settings