Imagine teaching one robot to pick up a soft object without crushing it. Today, that lesson may remain tied to the robot’s arm length, camera position, gripper and software. A second machine could require a new dataset and another round of training, even if it is being asked to perform essentially the same task.
“[2] . Our read The verdict is Overhyped , for the headline rather than the research. GR00T 2 looks like a real step in getting skills to move between robots, and Nvidia documents it more openly than most launches. But”
Nvidia is trying to change that pattern with GR00T 2, a new robotics model intended to learn physical skills from demonstrations and transfer them between different robot bodies.
In its November 13 post on X, Nvidia described GR00T 2 as its most advanced open “System One” model for robotics research. The company said the model learns from human demonstrations and predicts how its actions will change the physical world. Nvidia also said the resulting skills can carry across robot embodiments, reducing the need to teach every machine from scratch.
A model that predicts what happens next
Nvidia’s research page describes GR00T 2 as a “world-action model.” Rather than treating a robot’s movement as an isolated command, the system is designed to predict future video and robot actions together. That gives it a way to connect an instruction, an observed scene and the likely consequences of a movement.
For a robot, that could mean more than knowing how to move a gripper toward a cup. It could involve anticipating how the cup will shift, how much force is needed and what the scene may look like after contact. The approach is aimed at skills requiring dexterity and adaptation, particularly when objects, surfaces or robot bodies differ from the examples used during training.
The model learns from both human and robot demonstrations, according to Nvidia. Human video can show the structure of a task from a person’s point of view, while robot data can provide information about control, movement and physical interaction. Combining the two is intended to give the model a broader understanding of how actions unfold.
Nvidia’s November 9 announcement said GR00T 2 is an open world-action model built on the company’s Cosmos platform. The announcement positioned it as a tool for dexterous manipulation and wider robotics research, rather than as software for only one commercial machine.
That open positioning could matter to universities, robotics startups and manufacturers that cannot afford to build a separate foundation model for every platform. A shared model could give researchers a common starting point for testing manipulation, navigation and other physical capabilities.
The promise of reusable robot skills
The most consequential claim is not that GR00T 2 can perform a particular task. It is that the skill can survive a change in hardware.
A robot with a human-shaped body may approach a task differently from a compact factory arm or a machine with wheels and a single gripper. If a model can identify the goal of an action and adapt its movements to each body, training data could become more reusable. A demonstration collected on one platform might help another machine learn the same underlying behavior.
That would address one of robotics’ persistent bottlenecks. Physical data is expensive to collect because robots must operate in the real world, where failed actions can damage equipment, spill materials or create safety risks. Retraining for every new gripper, sensor arrangement or body shape makes deployment slower and more costly.
The GR00T 2 whitepaper details the architecture, training data, cross-embodiment experiments and evaluation methods behind the model. Nvidia reports 40% zero-shot success across trials on 50 tasks, with higher results after fine-tuning. The whitepaper also discusses limitations and open problems, suggesting that transfer is not automatic or universally reliable.
Why the headline number needs context
The reported results have already drawn criticism. In an analysis titled “Nvidia’s GR00T 2 ‘completed 40 of 50 tasks’. Here’s how it counted,” Samarth Urs argued that Nvidia’s headline figure could overstate the model’s performance.
Urs said the model succeeded on each of 40 of 50 tasks at least once, while the reported success rate across trials was 40%. The analysis also noted that the tasks were selected from skills represented in other-robot or human data. Urs described the launch claim as “overhyped.”
That criticism does not establish that GR00T 2 lacks value, but it shows why demonstrations and aggregate figures are not enough to judge a general-purpose robotics model. Researchers will need to know how tasks were selected, how many attempts were allowed, how much fine-tuning was required and how performance changes on unfamiliar objects and environments.
The available reaction cited for this story is critical, and no opposing independent assessment was provided. Nvidia’s own materials remain the main evidence for the model’s capabilities.
If GR00T 2 performs as promised beyond curated demonstrations, robot training could begin to resemble teaching a transferable skill rather than programming a single machine. A worker might demonstrate a process once, then adapt it across a fleet of robots. For now, however, the open model is best understood as a significant research direction, not proof that general-purpose physical intelligence has arrived.
Maya Lindqvist is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Maya Lindqvist's usual length and in Maya Lindqvist's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.