NVIDIA GR00T and the Rise of Physical AI Robots
A robot arm that welds car doors doesn’t need to understand what a door is. A humanoid that’s supposed to tidy a warehouse, adapt to a spill, and hand a box to a person does. That gap is why physical AI robots, humanoid systems paired with reasoning foundation models like NVIDIA’s GR00T, are pulling in more enterprise capital than almost anything else in AI right now. If you’re the one signing off on a robotics budget in 2026, the question has quietly changed. It’s no longer “which arm do we buy.” It’s “whose brain is running it.”
The Brain-Body Problem: Why Hardware Alone Never Worked
Industrial robots have been welding, painting, and palletizing for four decades. What they haven’t been able to do is generalize. A pick-and-place arm programmed for one bin geometry breaks the moment the bin changes. That’s the limit hard-coded automation always hit: every new task meant new code, new engineers, new downtime.
Between 2024 and 2026, that limit started to move. Vision-Language-Action (VLA) models, the same transformer architecture family behind large language models, got repurposed to output motor control instead of text. Instead of programming a robot for a task, you train a foundation model on a broad range of tasks and let it generalize the way GPT generalizes across writing styles. Industry shorthand calls this the split between “physical AI” (the hardware) and “cognitive AI” (the reasoning layer on top of it). The body was never the bottleneck. The brain was.
“Humanoid robots will bring physical AI to the world’s largest industries, opening a multitrillion-dollar economic opportunity.” Jensen Huang, Founder and CEO, NVIDIA, source: NVIDIA Newsroom
Worth remembering: Huang sells the platform this quote is describing. That doesn’t make him wrong, but it’s the kind of incentive an enterprise buyer should weigh before treating vendor keynotes as market research.
Inside NVIDIA’s GR00T Platform: From N1 to the Isaac Reference Robot
NVIDIA announced Project GR00T at GTC 2024 as a foundation-model initiative for humanoid robots, paired with its Jetson Thor compute platform and built alongside partners including Boston Dynamics, Figure AI, Agility Robotics, and Unitree.
The first real release, GR00T N1, runs on a dual-system architecture modeled loosely on human cognition: a fast “System 1” for reflexive motor actions, and a slower “System 2” for deliberate planning. 1X Technologies put N1 to work running autonomous tidying tasks on its NEO Gamma robot.
By GTC Taipei on May 31, 2026, NVIDIA had moved from software release to full reference hardware: the Isaac GR00T Reference Humanoid Robot, combining a Unitree H2 Plus chassis, five-fingered Sharpa hands, Jetson Thor compute, and the Isaac GR00T software stack. Launch partners include Ai2, ETH Zurich, Stanford’s Robotics Center, and UC San Diego’s Advanced Robotics and Controls Laboratory.
“The future of humanoids is about adaptability and learning. NVIDIA’s GR00T N1 provides a significant boost to robot reasoning and skills, advancing our mission of creating robots that are not just tools, but companions.” Bernt Børnich, CEO, 1X Technologies, source: NVIDIA Newsroom
Numbers worth a caveat: NVIDIA frequently cites a global labor shortage of more than 50 million people as the driver behind generalist robotics demand. That figure comes from NVIDIA itself, not an independent labor economist, so treat it as a vendor framing device rather than a settled statistic when you’re building an internal business case.
The Competitive Field: Helix, Gemini Robotics, and Physical Intelligence’s pi
NVIDIA isn’t operating alone in this space, and neither should your evaluation.
Figure AI’s Helix was the first VLA model to output full upper-body humanoid control, and the first to run simultaneously across two collaborating robots. Figure has since scaled BotQ, its manufacturing line, from one robot a day to one an hour, a 24x throughput jump in under 120 days, with more than 350 third-generation units shipped, according to Figure’s own technical disclosures.
Google DeepMind and Boston Dynamics partnered in January 2026 to run Gemini Robotics foundation models on Atlas, DeepMind’s contribution being the reasoning layer rather than the chassis. Boston Dynamics unveiled an electric Atlas at CES 2026 rated to lift up to 110 pounds, with 2026 fleet deployments already committed to Hyundai and Google DeepMind.
Physical Intelligence, founded out of Berkeley, Stanford, and Google DeepMind alumni, is building the “pi” model family and was reportedly in talks in March 2026 for funding that would value the company above $11 billion, roughly double its valuation four months earlier, per Bloomberg reporting via TechCrunch.
“Think of it like ChatGPT, but for robots.” Sergey Levine, Co-founder and Chief Scientist, Physical Intelligence, source: TechCrunch, March 2026
GR00T vs. Helix vs. pi: Quick Comparison
| Model | Maker | Architecture | Commercial stage |
|---|---|---|---|
| GR00T N1 / Isaac | NVIDIA | Dual-system (fast reflex + slow planning) | Platform, licensed to hardware partners |
| Helix | Figure AI | Single VLA, full upper-body output | In-house, running on Figure 03 units |
| pi (π) | Physical Intelligence | General-purpose VLA, hardware-agnostic | Pre-revenue, research-first |
Notice what’s missing from that table: nobody has independently benchmarked these three against each other on the same task, same hardware, same environment. Every performance claim you’ll read this year comes from the company that built the model.
The Proof Point: Amazon’s DeepFleet and the ROI Enterprises Can Actually See
Of everything in this space, Amazon’s deployment is the one with real operating data behind it rather than a keynote demo. On July 1, 2025, Amazon announced its one millionth deployed robot and introduced DeepFleet, a generative AI model trained on Amazon SageMaker that coordinates fleet movement across warehouse floors. Amazon reports DeepFleet improves robot travel efficiency by 10 percent, and the fleet now supports more than 75 percent of the company’s global deliveries across 300-plus fulfillment centers.
That 10 percent figure matters more than any humanoid headline in this piece, because it’s the rare number in this space that comes from a company measuring its own production operations, not a lab demo. For a deeper breakdown of what that ROI actually looks like line by line, see our companion piece on Amazon’s DeepFleet efficiency gains.
Related reading on NeuralWired:
• Amazon’s 1 Million Robots: The Real ROI Story (2026)
The Reality Check: China’s Rental Boom and the Autonomy Gap
Here’s where the hype meets the floor. China now has more than 153,000 robot rental businesses in operation, and AGIBOT’s SHAREBOT subsidiary projects that market could hit $1.5 billion by the end of 2026. But CNN’s on-the-ground reporting from inside a state-backed facility found more than 120 humanoids performing single repetitive tasks, sorting packages, scooping popcorn, changing diapers, each one guided in real time by a human operator holding a controller. Not autonomous. Remote-operated.
That’s the gap between the marketing language of “general intelligence” and what’s actually shipping today.
“The current reality is that beyond the hype and exotic expectations, humanoids are nowhere near a public debut, and the costs remain prohibitively high at $100,000 or more.” Tom Dotan, Technology Journalist, Newcomer
Retail humanoid pricing in China starts around $19,000 for entry models and climbs past $100,000 for advanced units, per CNN. And the long-range forecast getting quoted everywhere, Morgan Stanley’s estimate of one billion humanoids in use by 2050 in a market worth over $5 trillion, comes with a buried caveat: Morgan Stanley itself doesn’t expect adoption to accelerate for at least another decade. That’s a 2050 projection being used to justify 2026 budget requests. Read it that way.
What This Means for Your Procurement Roadmap
If you’re evaluating robotics deployment right now, three things follow from all of this:
- Evaluate the model roadmap, not just the spec sheet. Mechanical specs and MTBF used to be the whole conversation. Now you’re also underwriting a software vendor’s training data, update cadence, and long-term model support, exactly like choosing a cloud provider.
- Pilot in high-repetition environments first. Packaging, sorting, and fixed-station tasks are where DeepFleet-style gains are documented. General-purpose humanoid autonomy in unstructured environments is not there yet, whatever the demo reel shows.
- Price in vendor concentration risk. A small number of foundation-model providers (NVIDIA, Google DeepMind, Physical Intelligence) sit underneath most of the hardware brands. That’s a single point of failure the same way relying on one cloud region is.
Our read: the enterprises winning here in 2026 aren’t the ones with the flashiest humanoid pilot. They’re the ones treating this exactly like the early cloud migration decisions of a decade ago, boring, staged, and reversible.
Frequently Asked Questions
What is physical AI?
Physical AI refers to vision-language-action (VLA) models that let robots perceive, reason about, and act in the real world, combining the language-understanding approach of LLMs with real-time motor control, rather than the rigid, pre-programmed automation used in traditional industrial robotics.
What is NVIDIA GR00T?
GR00T is NVIDIA’s foundation-model platform for humanoid robots. It pairs open VLA models with the Jetson Thor computing chip to give robots generalized reasoning and manipulation skills, and is used by hardware partners including 1X Technologies, Figure AI, and Boston Dynamics.
How many robots does Amazon use?
As of its July 2025 announcement, Amazon operates more than one million robots across 300-plus fulfillment centers, coordinated by its DeepFleet AI model, supporting more than 75 percent of the company’s global deliveries.
Are humanoid robots actually being used by companies yet?
Yes, but in narrow, structured settings. Amazon uses them in logistics, BMW on factory floors, and Figure AI in warehouse pilots. Independent verification of general-purpose autonomy in unstructured environments, like a home or an unpredictable warehouse floor, remains limited.
How much do humanoid robots cost?
Pricing varies widely by capability. Entry-level Chinese humanoid models start around $19,000, while advanced industrial units can exceed $100,000. Figure and other Western manufacturers have not published consistent public pricing, so treat any specific figure outside verified company disclosures with caution.
Where This Goes Next
What you now know that a lot of the coverage on this topic skips: the robot body stopped being the interesting part around 2024. The real competition is happening one layer up, in the foundation models deciding what these machines do with their hands. NVIDIA, Figure AI, Google DeepMind, and Physical Intelligence are all racing to become the default cognitive layer for physical robots, the same land grab that happened with cloud infrastructure and, before that, mobile operating systems.
Over the next six to eighteen months, watch for three things: whether Figure’s manufacturing scale-up holds up under real warehouse conditions rather than staged livestreams, whether Physical Intelligence closes that funding round and what valuation it lands at, and whether any independent lab publishes a head-to-head benchmark of GR00T, Helix, and pi on identical hardware. That third one is the piece of evidence this whole market is currently missing.
Want the next update on this before it hits the wire? Subscribe to The Neural Loop at neuralwired.com/newsletter.