Menu

Nvidia GR00T 2 teaches robots a new task from five demonstrations: what it means for your business

Comingwave team · 7 minute read · published

A warehouse worker packs folded cloth into a carton while a tripod camera records his hands under warm orange light.

Nvidia GR00T 2 is a new AI model that teaches robots to handle objects, and Nvidia says it can pick up a new task from as few as five demonstrations. Nvidia introduced it on 9 November 2026 at the Conference on Robot Learning (CoRL) in Austin, Texas, alongside a new version of its robot training software and new open benchmarks for industrial work. For business owners, it is a clear sign of where robot automation is heading: less programming by specialists, more learning from people doing the job.

What happened

Nvidia's blog post of 9 November 2026, "Practice Makes Precise", says the company is introducing GR00T 2, Isaac Lab 3.0 and new open benchmarks at CoRL, the annual research conference on robots and machine learning. Nvidia's event page says CoRL 2026 runs from 9 to 11 November 2026, with workshops on 12 November.

The headline release is GR00T 2. Nvidia calls it an open frontier world action model built on NVIDIA Cosmos. Physical AI is the term for AI that controls machines acting in the real world, rather than writing text or images on a screen.

A world action model is different from the chat models most offices know. According to the GR00T 2 research page, it predicts future video and the robot's actions together. In plain terms, before the robot moves, the model imagines what the scene should look like a moment later, then produces the motor commands to get there. Nvidia says this lets it learn physical skills from video of people at work, not only from robots being steered by hand.

Nvidia says GR00T 2 was trained on robot and human data, including 100,000 hours of its open source human egocentric video, meaning footage filmed from the worker's own point of view. The model runs on parallel grippers, dexterous robot hands and humanoids.

Key details

ItemWhat Nvidia says
Announced9 November 2026, at CoRL in Austin, Texas
GR00T 2An open world action model on NVIDIA Cosmos, for grippers, dexterous hands and humanoids
Learning from few examplesOn 50 unseen tasks: 40% success with no demonstrations, 71% with one, 95% with five
Agent controlAI agents can call GR00T 2 through an interface, run skills in sequence and recover from errors
GR00T 2 TurboA faster version that runs on a Unitree G1 humanoid, 124 ms per prediction on its Jetson Thor computer
Isaac Lab 3.0Up to 2.5x the training throughput of Isaac Lab 2.3 in one dexterous manipulation benchmark
New benchmarksDexBench: 15 tasks across four scenes in retail, logistics, manufacturing and hospitality; VABAR: six industrial tasks
AvailabilityThe research page says "Code Available in December"; no licence terms or prices are given

The research page gives more detail on the results. The 50 tasks had never been practised on the target robot; they appeared only in human video or in other robots' data. With five demonstrations per task, success rose to 95% after a 30-minute fine-tune on a single RTX 5090 graphics card.

The agent results stand out. With a reasoning agent planning the steps and watching for mistakes, GR00T 2 assembled a toy plane, a toy tractor and a toy car in 98, 97 and 97 of 100 trials. The same model without the agent managed between 71 and 84. That is the pattern we see in office software too: one AI plans and checks, another does the work.

Nvidia names early users, including Foxconn for server manufacturing and Siemens across industrial use cases.

Why it matters

Until now, teaching an industrial robot a new job has usually meant programming every movement, or many hours of a person steering the robot by remote control. Both are slow and expensive, which is why robots mostly do one fixed task in a large factory. A model that learns from a few demonstrations, and from video of people, attacks that cost directly.

Simulation is the other half. Nvidia says Isaac Lab, working with the open source Newton physics engine, can now model soft and flexible objects such as cables and cloth. One example in the post, AlphaNIST, had a single policy learn all 20 tasks on the NIST Assembly Task Board, handling millimetre-scale parts. Nvidia reports more than 98% training success, against 81% for separate policies trained per task, with roughly 39x less GPU time.

The partner list shows where this work is aimed. Dyna Robotics is training its semi-humanoid Taku robot for warehouses, hotels and food services, and Walden Robotics is training a general-purpose robot for manufacturing and logistics. These are the same settings where Agility Robotics already runs its Digit humanoids, as we covered when Agility's Nasdaq listing was approved.

A word of caution, too. Every figure here comes from Nvidia's own lab tests and its partners' early trials. Nvidia says two of its papers are among six finalists for the conference's Best Paper Awards, which shows the research community takes the work seriously, but none of these numbers is an independent test in a working business.

What this means for businesses

For most small and medium businesses in Australia, GR00T 2 is not something to buy. It is a building block for robot makers. It matters for the robots those makers sell next: machines that can be shown a job rather than programmed for it.

The work it targets is easy to recognise: picking, sorting, packing, labelling, folding and simple assembly. If you run a warehouse, a commercial laundry, a kitchen or a small production line, those are the tasks to watch. Our page on technology for logistics and transport businesses covers the systems around that work, which any robot would need to connect to.

The agent side matters as much as the robot. GR00T 2 is designed to take instructions from an AI agent through an interface, much as office agents call tools today. The same questions apply: what is the agent allowed to do, who watches it, and how is it stopped. Our guide to permissions, data access and cost for AI agents is a good starting point.

A short checklist for the months ahead:

  • List the repetitive handling tasks in your business and roughly how many hours each takes a week.
  • Check that orders, stock and jobs live in connected systems, not on paper; a robot can only act on data it can reach.
  • Ask any equipment supplier who pitches AI robots which model they use, where its data goes, and what support looks like after installation.
  • Treat lab success rates as a starting point and ask for a trial on your own tasks before you sign anything.

Supply chains are where this lands first. If you import goods, from spices bought from an exporter such as Truly Ceylon to parts for a workshop, the warehouses that pack and ship them are among the first places these robots are being trialled. Nvidia also shares 100,000 hours of its human video openly on Hugging Face, the platform at the centre of Nvidia's Hugging Face deal.

Most of this preparation is not about robots at all: it is about connected systems and clean processes. Our digital transformation work includes a process audit and a staged roadmap. If you would like to talk through where automation fits your operation, ask us for a free first consultation and a written quote.

Frequently asked questions

What is Nvidia GR00T 2?

GR00T 2 is what Nvidia calls an open frontier world action model for robots, built on its Cosmos models. It predicts future video and robot actions together, learns from human video as well as robot data, and runs on robot grippers, dexterous hands and humanoids.

Can my business download GR00T 2 today?

Not yet in full. Nvidia's research page says "Code Available in December", and it does not give a licence or a price. Nvidia presents it to researchers and robot developers, so most businesses would meet it inside a robot or a system built by a supplier.

How many demonstrations does GR00T 2 need to learn a task?

On 50 tasks the target robot had never practised, Nvidia reports 40% success with no demonstrations, 71% with one demonstration per task and 95% with five, after a 30-minute fine-tune on a single RTX 5090. These are Nvidia's own lab results, not independent tests.

Which industries are testing GR00T 2?

Nvidia says Foxconn is applying it to server manufacturing, Siemens is evaluating it across industrial use cases, Nexuni is testing it for laundry handling and MicroAGI is using it to swap Ethernet cables with a robotic arm.

Should a small business plan to buy a robot because of this?

Not because of this announcement alone. It is a research and developer release, with no product, price or delivery date. The useful step is to map the repetitive handling work in your business and keep your systems and data ready for automation, whatever form it takes.

Key takeaways

  • Nvidia introduced GR00T 2, a world action model for robots, at CoRL on 9 November 2026.
  • Nvidia reports 95% success on 50 unseen tasks after five demonstrations each, in its own lab tests.
  • The model learns from video of people as well as robot data, and AI agents can direct it.
  • Code is due in December; there is no product, price or licence for businesses yet.
  • Prepare by mapping repetitive handling work and connecting your systems and data.

Need help with your business technology?

Tell us what you need. We aim to reply quickly.

Get a free quote