At Nvidia’s GPU Technology Conference in San Jose this March, the hallways weren’t just filled with software engineers and chip designers. They were filled with robots. Dozens of them — walking, rolling, gripping, stacking, and in at least one case, folding towels. The spectacle wasn’t a sideshow. It was the main event.
GTC 2025 marked the moment when humanoid robots and their less anthropomorphic cousins moved from research curiosities to something closer to commercial products. Nvidia CEO Jensen Huang used his keynote to announce a wave of new tools, platforms, and partnerships designed to accelerate the development of physical AI — machines that can perceive, reason about, and act in the real world. But the real story was on the show floor, where more than a dozen robotics companies demonstrated hardware that ranged from eerily lifelike humanoids to purpose-built industrial arms, all increasingly powered by Nvidia’s silicon and software stack.
CNET’s reporting from the event captured the breadth of what was on display. The coverage cataloged encounters with robots from companies spanning China, the United States, South Korea, and beyond — each with a distinct approach to the question of what robots should look like, what they should do, and how soon they’ll be doing it in your warehouse, your factory, or your home.
Start with the humanoids. They dominated the floor.
Figure, the Sunnyvale-based startup that has attracted billions in funding, showed its Figure 02 robot performing tasks that required not just dexterity but a form of spatial reasoning. The robot could pick up objects, place them precisely, and adapt to changes in its environment. Figure has been vocal about its ambitions to deploy humanoid robots in BMW manufacturing facilities, and the GTC demo suggested the company is making tangible progress toward that goal. The robot’s movements were fluid but deliberate, lacking the jerky uncertainty that characterized earlier generations of bipedal machines.
Then there was Unitree, the Chinese firm whose H1 humanoid has become something of an internet sensation for its backflipping videos. At GTC, Unitree showed a more practical side. Its robots demonstrated walking gaits that handled uneven surfaces and transitions between different floor types — the kind of mundane capability that matters enormously in real-world deployment. Unitree’s price point, reportedly a fraction of competitors like Boston Dynamics, has made it a favorite among researchers and startups looking for affordable humanoid platforms.
Agility Robotics brought Digit, its warehouse-focused bipedal robot that has already been piloted in Amazon fulfillment centers. Digit doesn’t try to look human. Its legs bend backward, ostrich-style, and its “face” is a simple sensor array. But it can pick up totes, carry them across a warehouse floor, and place them on shelves — repetitive tasks that are physically demanding for human workers and difficult for traditional automation to handle because of the variability involved. The company has been refining Digit’s ability to work alongside people without requiring the kind of safety caging that industrial robots typically need.
Not every robot at the show walked on two legs. Far from it.
Boston Dynamics, arguably the most recognized name in robotics, demonstrated its Spot quadruped and its Atlas platform. Atlas, which Boston Dynamics recently redesigned as a fully electric humanoid after years of hydraulic prototypes, moved with a smoothness that suggested significant engineering progress. Spot, meanwhile, continues to find commercial traction in industrial inspection, oil and gas facilities, and construction sites — environments where its four-legged stability gives it an advantage over wheeled or tracked robots.
The Chinese contingent was impossible to ignore. Companies like Agibot, Galbot, and XPENG Robotics all showcased humanoid or semi-humanoid platforms. XPENG, better known for its electric vehicles, has been pouring resources into its Iron humanoid robot, positioning it as a natural extension of the company’s expertise in batteries, motors, and AI-driven control systems. The crossover between EV engineering and humanoid robotics is becoming a recurring theme: the actuators that drive a robot’s joints share fundamental engineering DNA with the motors that drive electric car wheels.
Agibot’s presentation focused on dexterous manipulation — the ability to handle objects of varying shapes, sizes, and materials without crushing them or dropping them. This is one of the hardest unsolved problems in robotics. A human hand performs thousands of micro-adjustments when picking up a coffee cup versus a raw egg versus a sheet of paper. Teaching a robot hand to do the same requires not just sophisticated hardware but massive amounts of training data, which is where Nvidia’s simulation platforms come in.
And that’s really the thread connecting all of these machines: Nvidia’s growing role as the infrastructure provider for the entire physical AI sector.
Huang announced several new offerings at GTC aimed squarely at robotics developers. Newton, a physics engine designed specifically for robotics simulation, promises to let developers train robots in virtual environments that accurately replicate real-world physics — friction, gravity, deformation, fluid dynamics. The idea is straightforward but extraordinarily difficult to execute well. If a simulated environment doesn’t match reality closely enough, a robot trained in simulation will fail when transferred to the physical world, a problem researchers call the sim-to-real gap.
Nvidia also introduced GR00T N1, an open foundation model for humanoid robots. Foundation models have transformed natural language processing and image generation over the past three years. Nvidia is betting the same approach — training large neural networks on vast datasets to produce general-purpose capabilities that can be fine-tuned for specific tasks — will work for robotics. GR00T N1 is designed to give humanoid robots a baseline understanding of how to move through and interact with the physical world, which developers can then specialize for particular applications.
The company’s Omniverse platform, which provides tools for building and running digital twins of physical environments, has become central to how many robotics companies develop and test their systems. A warehouse operator can build a virtual replica of its facility, populate it with virtual robots, and run thousands of hours of simulated operations before a single physical robot rolls onto the floor. This approach dramatically reduces development time and, critically, the risk of expensive hardware damage during testing.
Nvidia’s Jetson Thor computing platform, designed specifically for humanoid robots, provides the onboard processing power these machines need to run AI models in real time. A humanoid robot operating in an unstructured environment — a home, a hospital, a construction site — can’t afford the latency of sending data to the cloud for processing and waiting for instructions to come back. It needs to make decisions locally, instantly. Jetson Thor is Nvidia’s answer to that requirement.
So where does all of this actually lead?
The near-term applications are industrial. Warehouses, factories, logistics hubs. These are environments with enough economic value to justify the current cost of robotic systems and enough structure to make the AI problem tractable, even if not fully solved. Amazon’s experiments with Agility’s Digit, BMW’s partnership with Figure, and numerous deployments of Boston Dynamics’ Spot all point in this direction.
The medium-term opportunity, and the one generating the most investor excitement, is broader commercial deployment. Robots that can operate in retail environments, healthcare facilities, agriculture, and eventually homes. The laundry-folding demo at GTC — performed by a robot from a company called 1X, which has backing from OpenAI — got enormous attention precisely because it represents a task that’s universally understood and universally disliked. If a robot can fold your laundry reliably and affordably, the consumer market opens wide.
But reliability and affordability remain the hard parts. The robots at GTC were impressive in controlled demo environments. The gap between a polished conference demonstration and a product that works eight hours a day, five days a week, in a messy real-world setting is still substantial. Motors wear out. Sensors get dirty. Software encounters edge cases that weren’t represented in training data. Every robotics company at GTC knows this intimately.
The investment flowing into the sector is staggering. Figure has raised over $2.6 billion. Agility Robotics has secured hundreds of millions. Chinese competitors are backed by some of the largest technology conglomerates on the planet. Nvidia itself is making enormous capital commitments to robotics-related R&D, viewing physical AI as one of the primary growth vectors for its GPU and software businesses as the data center AI buildout begins to mature.
There’s a geopolitical dimension too. The United States and China are both treating humanoid robotics as a strategic technology, much as they’ve treated semiconductor manufacturing and artificial intelligence. China’s Ministry of Industry and Information Technology has published explicit goals for humanoid robot development, and Chinese companies are moving fast. The GTC show floor reflected this competition in miniature — American and Chinese robots standing side by side, each representing billions in national investment and ambition.
One thing that struck observers at the event was how quickly the conversation has shifted from “if” to “when.” Two years ago, humanoid robots were widely viewed as science fiction vanity projects — impressive but impractical. The combination of better AI models, cheaper and more capable hardware, improved simulation tools, and massive capital investment has compressed timelines dramatically. No serious industry analyst is predicting humanoid robots in every home by 2027. But deployment in structured commercial environments within the next two to three years? That’s no longer a fringe position.
The companies that succeed will be the ones that solve the integration problem — not just building a robot that can perform a task in isolation, but building a robot that can operate within existing workflows, alongside human workers, using existing infrastructure. That means software that’s reliable, hardware that’s maintainable, and business models that make economic sense for customers. The flashiest demo doesn’t always win. Sometimes the robot that quietly moves totes from point A to point B, eight hours straight, without breaking down, is the one that changes an industry.
Nvidia is positioning itself to profit regardless of which robot company ultimately dominates. By providing the chips, the simulation tools, the foundation models, and the development platforms, the company is building the equivalent of what Intel and Microsoft built for the PC era — the essential layer that everyone else builds on top of. It’s a familiar playbook, executed at a moment when the market for physical AI is just beginning to take shape.
GTC 2025 wasn’t the moment robots arrived. It was the moment the industry stopped debating whether they would and started arguing about how fast.
WebProNews is an iEntry Publication