GR00T 2: Robots That Learn by Watching You

Isometric illustration in coral, slate blue and mint showing a robot learning dexterous tasks by watching human video.

NVIDIA introduced GR00T 2 on 9 November 2026 at the Conference on Robot Learning, known as CoRL, in Austin, Texas. It is a robot model that learns dexterous manipulation — skilled hand work — largely by watching first-person video of humans, not just by copying robot demonstrations.

That is a big shift. Robots have usually learned by watching other robots, which is slow and expensive. Humans generate enormous amounts of video of everyday tasks, and GR00T 2 is built to learn from that video and transfer it to a robot body.

What GR00T 2 actually is

GR00T 2 is what NVIDIA calls an "open frontier world action model", built on NVIDIA Cosmos. A world action model predicts future video and robot actions together, in a single diffusion transformer. In plain terms, it reads recent video, the instruction and the robot's state, then generates a short chunk of future video and motor commands.

It is built on Cosmos-Predict2.5-2B, a 2-billion-parameter transformer. One checkpoint covers parallel-gripper arms, dexterous hands and Unitree's G1 humanoid. It succeeds the GR00T N1 series (N1.6 and N1.7).

Learning from human video

The key idea is the training mix. NVIDIA says GR00T 2 was pretrained on 402,730 hours of first-person human video, alongside 10,184 hours of handheld gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation — a human remotely controlling a robot to show it a task.

The human video is the largest part by far. First-person video shows hands picking things up, opening containers, folding and assembling — skills hard to programme by hand.

NVIDIA is releasing 100,000 hours of that human egocentric data, called Human-1M-100K, as open source under a CC BY 4.0 licence, including video, language annotations, and head and hand motion data.

How well does it work?

NVIDIA reports its own results, and they have not been independently verified. On 50 unseen tasks with no demonstrations, NVIDIA says GR00T 2 averaged 40 percent success — it completed 40 of the 50 tasks at least once. With one demonstration per task, success rose to 71 percent; with five, it reached 95 percent.

Zero demonstrations is the hardest test, because the robot has never seen that exact task before. NVIDIA says success on new tasks in unfamiliar environments is double that of previous leading models — a company claim, not independently verified. It also ranks among the top three models on RoboCasa365, RoboLab, RoboDojo and MolmoSpaces, NVIDIA says.

Isaac Lab 3.0 and the wider launch

GR00T 2 did not arrive alone. NVIDIA also highlighted Isaac Lab 3.0, its open-source robot-learning framework, generally available since 20 October 2026. Isaac Lab lets AI agents set up training environments, switch between PhysX and Newton physics engines, and apply a new reinforcement post-training library to models like GR00T.

NVIDIA says Isaac Lab 3.0 delivers up to 2.5 times the training throughput of Isaac Lab 2.3 on a dexterous manipulation benchmark — NVIDIA's own figure, not an independent test. Also at CoRL, NVIDIA introduced DexBench and VABAR benchmarks for testing robot policies in simulation, and an "agentic robotics" approach with UC Berkeley.

Who is already testing it

NVIDIA named several early users, all testing or evaluating the model — no production rollout has been announced. Nexuni is testing robotic laundry handling, Foxconn is applying it to server-manufacturing automation, and Siemens is evaluating industrial use cases. MicroAGI is using GR00T 2 post-trained with reinforcement learning for Ethernet cable swapping with a YAM robotic arm.

NVIDIA calls GR00T 2 open, but its research page says code will be available in December. The data release is open source under a CC BY 4.0 licence, but the code itself is not yet out.

Key takeaways

  • NVIDIA introduced GR00T 2 on 9 Nov 2026 at CoRL in Austin, built on Cosmos.
  • It predicts future video and robot actions together.
  • It learns largely from 402,730 hours of first-person human video.
  • NVIDIA is releasing 100,000 hours of that data as open source (CC BY 4.0).
  • NVIDIA reports 40% success with no demonstrations, 71% with one, 95% with five.
  • Isaac Lab 3.0 is available since 20 Oct 2026, with up to 2.5x training throughput claimed.
  • Early users are testing it for laundry, server manufacturing and cable swapping.
  • Code is promised for December.

Conclusion

GR00T 2 bets that the fastest way to teach robots skilled hand work is to let them watch humans, at a scale no robot-only dataset could match. The jump from 40 percent with no demonstration to 95 percent with five suggests a little guidance goes a long way.

The limits are clear: the results are NVIDIA's own, the code is not out until December, and early users are still testing.

FAQ

How does GR00T 2 learn from human video?

It is pretrained largely on first-person human video — 402,730 hours — alongside gripper demonstrations and robot teleoperation. It reads recent video, the instruction and the robot's state, then generates future video and motor commands.

How well does GR00T 2 do on tasks it has never seen?

NVIDIA reports 40 percent average success on 50 unseen tasks with no demonstrations, rising to 71 percent with one demonstration and 95 percent with five. These are NVIDIA's own figures, not independently verified.

When can I try GR00T 2 myself?

NVIDIA calls it open, but its research page says code will be available in December. The 100,000-hour Human-1M-100K dataset is being released as open source under a CC BY 4.0 licence.

Sources

Related reading on Daily Tech & AI:

Researched with AI assistance and reviewed for accuracy.

Follow on WhatsApp

Follow Daily Tech & AI on WhatsApp for the latest tech and AI updates: https://www.whatsapp.com/channel/0029Vb9XzdHDJ6H4TFNAl61r

Comments