Table of Contents
- Why This Robot News Feels Different?
- What Exactly Is NVIDIA Isaac GR00T?
- The Timeline: From June to This Week
- Meet the Agile ONE SN Reference Humanoid
- The Brain Inside: Jetson Thor and the GR00T Models
- The Unitree H2 Plus Reference Robot and Its Hands
- Why a Reference Robot Changes How Humanoids Learn
- Who Is Building on Isaac GR00T Right Now
- Release Date and Price: Can You Buy One?
- Comparison Table: Two Reference Humanoids Side by Side
- Safety, Jobs and Power Bills: The Honest Concerns
- Frequently Asked Questions
- Final Thoughts
1. Why This Robot News Feels Different?
Most robot headlines follow the same script. A shiny machine walks across a stage, someone claps, and the video goes viral for a day.
This week’s news is not that.
There is no single robot to point at and say “that is the one.” Instead, NVIDIA is handing the robotics world a starter kit, and the whole industry is lining up to use it.
That is why the NVIDIA Isaac GR00T story matters so much right now. It is less about one humanoid and more about how every humanoid might learn from here on.
If you want the other big robot story from the same week, read our full breakdown of the NVIDIA Microduck robot shown at GTC Berlin before you continue.
In this guide, I will keep things simple. No heavy jargon, no hype for the sake of hype. Just what was announced, who is involved, what is still unknown, and why you should care.
Let us start with the basics.
2. What Exactly Is NVIDIA Isaac GR00T?
Here is the part that confuses most people.
Isaac GR00T is not one robot. It is a platform.
Think of it like a phone operating system. Android is not a phone. It is the software that lets many different phones work. Isaac GR00T plays a similar role for humanoid robots.
According to NVIDIA’s own description, the platform covers the whole journey, from data capture and generation to robot model evaluation and deployment, and it is built on open software and models.
Here is what sits inside the Isaac GR00T family, in plain words:
- Robot foundation models. These are the “brains” that learn general skills instead of one narrow trick.
- Simulation tools. Robots can practice millions of times in a virtual world before touching a real floor.
- Data pipelines. Tools that help developers create and organize the training data robots need.
- Deployment software. The pieces that move a trained skill onto a real machine.
The models are the headline act. NVIDIA announced that GR00T N1.7 is available in early access with commercial licensing, which brings dexterous control and generalized skills toward real production use.
So when you read “Isaac GR00T humanoid robot” in a headline, remember this. The robot is the body. GR00T is the learning system that makes the body useful.
3. The Timeline: From June to This Week
The story did not start this week. It has been building for months, and it helps to see the steps in order.
| Date | What Happened | Why It Matters |
|---|---|---|
| Earlier in 2026 (GTC) | GR00T N1.7 opened in early access with commercial licensing, and GR00T N2 was previewed | Showed that the models were moving toward real products |
| June 1, 2026 | The Isaac GR00T Reference Humanoid Robot was announced at GTC Taipei | First open reference design built on Jetson Thor |
| October 21, 2026 | The Isaac platform expanded with the Agile ONE SN reference humanoid, new datasets and a runtime | A second reference body, aimed at research labs |
| October 26, 2026 | Agile Robots shared production plans of up to 2,000 humanoids | Moves the idea from lab demo toward real units |
The first big moment came in June. NVIDIA announced the Isaac GR00T Reference Humanoid Robot at GTC Taipei, and it was described as the first open humanoid reference design built on Jetson Thor and the Isaac GR00T open software stack.
The second big moment landed on October 21, during the Berlin event week. That is where Agile ONE SN entered the story.
NVIDIA has been pushing hard on robots all year, and our list of mind boggling NVIDIA inventions shows how this humanoid push fits into the bigger company picture.
Notice the pattern here. NVIDIA is not trying to build and sell a humanoid to you. It is building the tools and the reference designs so that other companies can.
That is a very different strategy, and a very smart one.
4. Meet the Agile ONE SN Reference Humanoid
Now for the star of this week.
On October 21, NVIDIA announced a major expansion of Isaac, and one of its headline pieces was Agile ONE SN, described as a humanoid reference platform built on Isaac GR00T and Jetson Thor, with planned integration of the NVIDIA Halos safety stack.
Let me break that sentence down so it makes sense.
- Agile ONE SN is the robot, made by a company called Agile Robots.
- Isaac GR00T is the learning software that comes with it.
- Jetson Thor is the powerful computer chip that runs inside the robot.
- Halos is NVIDIA’s safety system, which is planned for later integration, so it is not a finished feature yet.
The goal is simple. Researchers should not have to spend months stitching parts together before they can even start testing ideas.
The Agile Robots founder and CEO, Zhaopeng Chen, framed the idea as usefulness at scale, meaning one adaptable machine that learns many skills and works in spaces built for people.
I like that framing because it is honest. The big challenge in robotics is not making a robot walk. It is making one robot useful in many places.
The Numbers Everyone Is Asking About
The most interesting update came just two days ago. Agile Robots told Reuters it plans to build between 1,000 and 2,000 compact humanoid robots next year, with production of the Agile ONE S and ONE SN models scheduled to begin by the end of 2026.
Two details matter here:
- Who it is for. Stanford University and ETH Zurich were named as potential customers, so this is aimed at research labs first.
- When. The models are expected to become available in early 2027.
So no, you will not see one in a shop window next month.
But a thousand or more research humanoids, all running the same open learning stack, could speed up robot progress faster than any single flashy demo.
5. The Brain Inside: Jetson Thor and the GR00T Models
A robot body is just metal and motors until something smart sits inside it. That something has two parts, and it helps to keep them separate in your head.
Part One: The Computer (Jetson Thor)
Jetson Thor is the onboard computer. It lives inside the robot, so the machine can think and react without sending every decision to a faraway server.
Why does that matter? Imagine a humanoid carrying a box across a room. If it had to wait for a cloud connection before every step, one bad Wi-Fi moment could mean a dropped box or worse. Local compute keeps the reaction fast.
NVIDIA says Thor handles the advanced reasoning and control for the robot, and the reference designs are built around it.
Part Two: The Learning Software (GR00T Models)
The second part is the GR00T family of robot foundation models. A foundation model is a big, general-purpose AI trained on huge amounts of data, so it can pick up many tasks instead of one.
Here is a simple way to picture it:
- Old robots: one program, one job. Move this arm from here to there, forever.
- GR00T style robots: one general brain that can be taught new skills from examples, simulation and instructions.
Two versions are worth knowing:
| Model | Status | What It Means for You |
|---|---|---|
| GR00T N1.7 | Early access with commercial licensing | Companies can already build real products on it |
| GR00T N2 | Previewed by Jensen Huang, based on DreamZero research | The next generation, still a preview, so no final release details yet |
I want to be careful on that last row. N2 was shown as a preview only. Anyone telling you it is “out now” is guessing.
One independent write-up published this week describes the platform as a full AI brain and development ecosystem rather than just a humanoid robot, covering things like foundation models, computer vision, simulation and robot learning. That matches how NVIDIA itself positions it.
How a Robot Brain Learns to Act?
Here is the bit most people skip, and it is the most interesting. Teaching a robot is a lot like teaching software agents to use apps. You show the system a goal, let it try, correct mistakes, and repeat until the skill sticks.
The same idea is already changing screens before it changes factories, which is why our guide on AI agents that click buttons for you is a useful warm up for understanding robot learning.
The difference is the stakes. When a software agent clicks the wrong button, you get a bad email. When a six foot robot makes the wrong move, someone could get hurt. Hold that thought, because we will come back to it in the safety section.
6. The Unitree H2 Plus Reference Robot and Its Hands
Before Agile ONE SN, the first reference body was the one announced in June. It deserves its own section, because it shows exactly what NVIDIA means by “open.”
What Is It Made Of?
The Isaac GR00T Reference Humanoid Robot combines four main parts:
- A Unitree H2 Plus humanoid body
- Sharpa five fingered hands built for delicate, dexterous handling
- Jetson Thor as the onboard computer
- Isaac GR00T open software and models running the show
The hands are the quiet hero here. Walking gets the cheering, but real work happens with fingers. Opening a drawer, holding a screw or passing a cup of tea all need fine control.
A tech news report on the launch also notes the robot includes a remote emergency stop so researchers can shut it down instantly while testing new behaviors on a machine that stands about six feet tall.
I really like that detail. It is not glamorous, but it is exactly what a responsible lab needs.
Who Gets to Use It?
NVIDIA named four academic groups that will use the design to push humanoid research forward:
- Ai2
- ETH Zurich
- Stanford Robotics Center
- UC San Diego’s Advanced Robotics and Controls Laboratory
The point is access. Until now, a university that wanted a capable humanoid usually had to buy a closed system or build everything from scratch. An open reference design lowers that wall.
There is also a software side. NVIDIA said a reference workflow for the Unitree G1 robot is expected to arrive on GitHub and Hugging Face for developers, so people without the newest hardware can still join in.
7. Why a Reference Robot Changes How Humanoids Learn?
This is the heart of the whole story, so let me slow down.
What “Reference” Actually Means?
A reference design is a proven blueprint. It says, “Here is a working combination of body, brain and software. Start here.”
Think of how a standard laptop design helped hundreds of companies build laptops quickly. Nobody had to reinvent the screen hinge. The same logic applies to humanoids.
Three Big Changes This Could Bring
1. Faster starts. Teams skip months of basic setup and go straight to testing ideas.
2. Shared skills. If many labs use the same stack, a skill taught in one place has a better chance of working in another. That is huge for progress.
3. Fair competition. Smaller teams get tools that used to belong only to giant companies.
The Simulation Advantage
Real robots are slow and expensive to train. A fall can break a part, and every test takes real time.
Simulation fixes that. In a virtual world, a robot can practice a task thousands of times at once, then carry what it learned into the real machine.
NVIDIA’s October 21 package leaned hard into this idea. According to an industry news recap of the launch, the new tools help developers generate training data, scale robot learning, build simulation environments and deploy robot applications, which shortens the time from idea to working robot.
A few examples from the launch show the pattern:
- Flexion is using Isaac Lab and Newton to scale learning for Reflect, a unified robot model that combines reasoning with whole body movement for complex manipulation.
- Schaeffler is contributing its actuator technology as a foundational component for Newton, which helps simulate humanoid movement more realistically.
- Open robot learning workflows running on DGX Station are being backed by names such as 1X, Agile Robots, Humanoid, Maven Robotics, Mentee, MicroAGI, Rhoda, Sereact and Skild AI.
In short, the whole chain is getting connected. Data, simulation, training and deployment now live in one open toolbox.
8. Who Is Building on Isaac GR00T Right Now?
A platform is only as strong as the crowd using it. So who is actually in the room?
Humanoid Makers Adopting GR00T N Models
NVIDIA said AGIBOT, Humanoid, LG Electronics, NEURA Robotics and Noble Machines are adopting Isaac GR00T N models to speed up industrial deployments of their humanoids.
That list matters because it mixes startups with a household electronics giant. When a company like LG joins, it signals that this is not just a university experiment.
The Wider Physical AI Circle
The ecosystem goes well beyond humanoids. NVIDIA’s partner announcement named robot brain developers, industrial robot giants, surgical robot makers and humanoid pioneers, including:
| Group | Examples Named |
|---|---|
| Industrial robot makers | ABB Robotics, FANUC, KUKA, Universal Robots, YASKAWA |
| Humanoid builders | AGIBOT, Agility, Figure |
| Robot brain and AI teams | Skild AI, World Labs |
| Surgical and medical robotics | CMR Surgical, Medtronic |
| Other partners | Hexagon Robotics |
Notice something here. Many of these companies compete with each other. Yet they are building on the same NVIDIA foundation.
That is the real power move. NVIDIA does not need to win the humanoid race. It just needs every runner to wear its shoes.
A Quick Reality Check
I want to be fair to you as a reader. “Building on NVIDIA technology” can mean many things. For one company it might mean a full product. For another it might mean testing a single tool.
So read partner lists as a sign of momentum, not proof that thousands of finished humanoids are working in warehouses today.
9. Release Date and Price: Can You Buy One?
This is the question every reader asks, so let me answer it as honestly as I can.
Short answer: not like a phone or a laptop. Right now, these machines are aimed at researchers and developers, not shoppers.
Here is what we know:
- Unitree H2 Plus reference robot: NVIDIA said it would come from Unitree in late 2026. Pricing has not been disclosed.
- Agile ONE SN: production is planned to start by the end of 2026, with availability expected in early 2027.
- Agile Robots volume plan: between 1,000 and 2,000 units next year.
And here is what we do not know:
- The final price of either robot
- Whether regular businesses, and not only universities, can order one
- How soon the Halos safety integration will be ready on the Agile ONE SN
I am not going to guess a price for you. Any number you see online right now is speculation.
So if you are a student, a hobbyist or a curious reader, your best move is simple. Watch the software side. Open tools on GitHub and Hugging Face are usually where regular people can start touching this technology first.
10. Comparison Table: Two Reference Humanoids Side by Side
Here is a clean look at the two reference designs built on Isaac GR00T.
| Feature | Isaac GR00T Reference Humanoid (Unitree H2 Plus) | Agile ONE SN (Agile Robots) |
|---|---|---|
| Announced | June 1, 2026 at GTC Taipei | October 21, 2026 |
| Maker of the body | Unitree | Agile Robots |
| Onboard computer | Jetson Thor | Jetson Thor |
| Learning software | Isaac GR00T open models | Isaac GR00T |
| Hands | Sharpa five fingered hands | Not detailed in the sources I found |
| Safety stack | Remote emergency stop mentioned | Halos integration planned |
| Named research users | Ai2, ETH Zurich, Stanford Robotics Center, UC San Diego | Stanford and ETH Zurich named as potential customers |
| Availability | Late 2026 from Unitree | Early 2027 |
| Price | Not disclosed | Not disclosed |
Look at the two columns together. The bodies are different, but the brain and the learning tools are the same family.
That is exactly the point of a platform. The robot can change. The learning system stays familiar.
11. Safety, Jobs and Power Bills: The Honest Concerns
I would not be doing my job if I only gave you the exciting side. Humanoid robots raise real questions, and you deserve straight answers.
Concern 1: Safety
A six foot machine with strong motors and smart software needs strict guardrails.
That is why the planned Halos safety stack on Agile ONE SN matters. It is still a plan, not a finished feature, so do not treat it as proof that these robots are already safe in every setting.
It also raises a bigger question. What happens when an AI system with real freedom does something nobody expected?
We have already seen software systems behave in surprising ways, as our report on an AI agent that escaped its sandbox explains, and those lessons matter even more once the system has a physical body.
Some experts go further and worry about what physical AI could mean in the long run.
If you want the bigger picture of those fears, our breakdown of warnings about superintelligence, robots and bio risk covers what the loudest voices are saying and where they disagree.
My take is simple. Open platforms help because many eyes can test them. But openness alone is not safety. Testing, limits and emergency stops are what keep people protected.
Concern 2: Jobs
Whenever a humanoid robot makes news, one question follows. Will it take my job?
Honest answer: not tomorrow, and not in the same way for everyone. A research robot that arrives in early 2027 is not walking into your workplace next spring.
But the direction is worth watching. If one general robot can learn many skills, the old idea that robots only do one boring task starts to fade.
The pressure is already showing up in white collar work, which is the reason we mapped out the high paying jobs AI will erase first in a separate guide.
Younger workers feel it early, too.
If you are just starting your career, our look at how Gen Z entry level jobs are changing because of AI shows what is happening on the ground.
The useful takeaway is not panic. It is preparation. People who learn to work with these systems will have more options than people who ignore them.
Concern 3: Power Bills and Compute
Training robot brains takes serious computing. Simulation, data generation and model training all burn energy, and that energy has to come from somewhere.
This is part of a much larger debate about AI data center power demand and your electricity bill, and robotics will only add more weight to that conversation.
I do not have a figure for how much power this specific platform uses, so I will not invent one. What I can say is that the compute cost is real, and it is a fair thing to ask companies about.
12. Frequently Asked Questions
What is NVIDIA Isaac GR00T?
It is an open platform for humanoid robots. It includes robot foundation models, simulation tools, data pipelines and deployment software, so developers can train and run robots faster.
What is Agile ONE SN?
It is a humanoid reference platform from Agile Robots, built on Isaac GR00T and Jetson Thor. It is aimed at research labs, with production planned from the end of 2026 and availability expected in early 2027.
Is this the first NVIDIA humanoid robot?
Not exactly. NVIDIA does not sell its own humanoid. The first open reference design was announced in June 2026 and uses a Unitree H2 Plus body. Agile ONE SN is a second reference platform.
Can I buy an Isaac GR00T humanoid robot?
Not as a regular consumer product yet. These robots target researchers and developers, and prices have not been disclosed.
What is GR00T N2?
It is a next generation robot foundation model that Jensen Huang previewed at GTC. It was shown as a preview, so final release details are not confirmed.
Is Jetson Thor the robot?
No. Jetson Thor is the onboard computer that runs inside the robot. The body, the hands and the GR00T software are separate pieces.
13. Final Thoughts
Let me bring it all together.
The biggest lesson from this week is that the real action is not one robot. It is the shared foundation underneath many robots.
NVIDIA is building the road, and humanoid makers are lining up to drive on it. The Agile ONE SN is the newest car on that road, and a thousand or more could be built next year.
Will it change how every humanoid learns? It has a real shot, mostly because open tools spread fast. But the safety, jobs and power questions are still open, and they deserve your attention.
My advice is simple. Follow the software, watch the early 2027 availability news, and treat any wild claim about “robots everywhere tomorrow” with a smile and a pinch of salt.
I will keep tracking every update as it lands.
Related Reads You Should Not Miss
Also read: the numbers behind the AI bubble for OpenAI, Anthropic and SpaceXAI
Also read: what big tech CEOs are warning about AI and jobs
Also read: the full list of AI 2027 predictions
Also read: how AI in Hollywood could replace actors, writers and VFX by 2027