GR00T 2 explained: Nvidia's open world action model for robots

By LLMTest Lab · Nov 9, 2026 · 3 min read releaseroboticsopen-weight
On this page

On this page

  1. What is new
  2. What Nvidia reports
  3. Who should care
  4. How to follow up

Not tested by us yet. GR00T 2 controls robot bodies, and our harness only calls text and image models through an API, so we cannot run it. No hosted endpoint or download date has been published either. Everything below comes from Nvidia's CoRL robotics post and secondary coverage by The Next Web. We will update this post when the weights or an API are public.

GR00T 2 is an open "world action model" for humanoid robots, announced by Nvidia on November 9, 2026 at the Conference on Robot Learning in Austin. Nvidia says it completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations. If you work on robot manipulation, it is worth tracking. If you pick chat or coding models, it does not change your shortlist.

GR00T 2
Maker Nvidia
Announced 2026-11-09
Type World action model for humanoid robots, built on Nvidia Cosmos
Training data Robot data plus human video, including 100,000 hours of first-person footage Nvidia is releasing
Parameters Not published
License Not published
Availability Described as open; download location and date not stated
Hosted API None announced

What is new

Nvidia calls GR00T 2 an "open frontier world action model." The "world" part means the model predicts how its own actions will change what the robot sees before it moves. Earlier GR00T releases, such as N1.6 and N1.7, were vision-language-action models. According to Crypto Briefing, those used a 3B-parameter open model on a Cosmos backend. Nvidia has not said how large GR00T 2 is, so a size comparison with its predecessor is not possible yet.

The training recipe leans on human video. Nvidia says GR00T 2 learns skills from human demonstrations and transfers them between different robot bodies without robot demonstrations. It is also releasing the 100,000 hours of egocentric human footage it used, which is useful to other labs regardless of what happens to the model.

AI agents can drive it too. Nvidia says an agent can use the model's API and tools to run skills in sequence and recover from errors. The announcement gives no latency or reliability figures for that.

What Nvidia reports

These results are Nvidia's own, and nobody independent has reproduced them.

  • No task-specific training: 40 of 50 evaluated tasks completed at least once.
  • One demonstration per task: 71% average success across the 50 tasks.
  • Five demonstrations per task: 95% average success.
  • Rankings: top three in its categories on RoboCasa365, RoboLab, RoboDojo and MolmoSpaces, according to Nvidia.

Read the first number carefully. "At least once" is a much lower bar than reliable success, and The Next Web reports that Nvidia's own paper notes the tasks were picked from ones the model could already learn well. The 71% and 95% figures are the stronger evidence, but they come from a task set Nvidia chose and evaluated itself.

Nvidia also launched Isaac Lab 3.0 with the model, along with partner-built benchmarks called DexBench and VABAR. Those are the pieces to watch, because outside teams can run them and publish numbers that do not depend on Nvidia's setup.

Who should care

Robotics teams are the obvious audience, especially anyone who wants to train from human video instead of collecting expensive teleoperation data. Nvidia names Nexuni (laundry handling), Foxconn (server manufacturing), Siemens (industrial evaluation) and MicroAGI (Ethernet cable swapping) as early users. None of them has said it will use GR00T 2 in production.

If you are choosing language models, the only link is indirect: a robot that takes natural-language instructions still needs a planner, and an agent API on top of the model suggests Nvidia expects an LLM to sit in that seat. We have not seen what that integration looks like.

Wait if you need a license you can plan around. Without parameter count, license terms or a download date, you cannot budget hardware or check that your product can legally ship on it.

How to follow up

Check the Nvidia research page linked from the post for the weights, and look for third-party runs on DexBench and VABAR before trusting the headline numbers. To build your own comparison set once a model is available, see our guide on how to choose an LLM, and for another recent open release from a lab outside the usual names, read our Arete-27B explainer.

We will update this post once GR00T 2 is downloadable or independent results appear.

Related articles

Arete-27B explained: Nous Research's re-aligned Qwen3.8 27B
Arete-27B is Nous Research's Apache-2.0 fine-tune of Qwen3.8-27B, built to cut state-propaganda answers. Specs, vendor-reported evals, caveats, who should care.
How to choose an LLM in 2026: the definitive guide
A 7-step framework for picking the right LLM for any job. Real constraints, real benchmarks, real routing. Stop guessing from leaderboards.