NVIDIA has introduced GR00T 2, a robot foundation model that learns dexterous skills by watching people. Announced on Monday 9 November at the Conference on Robot Learning (CoRL) in Austin, Texas, it is trained on robot data and 100,000 hours of first-person human video — and NVIDIA says it can pick up a new task from as little as a single demonstration.
The launch attacks the biggest bottleneck in robotics: data. Teaching a robot a new skill usually means a human operator steering it through the task by hand. GR00T 2 instead learns from ordinary video of people at work and transfers those skills to robot bodies that have never practised the task.
What GR00T 2 is
NVIDIA calls GR00T 2 an open frontier “world action model”, built on its Cosmos world-model platform. Unlike a conventional robot policy, which maps camera views straight to motor commands, a world action model predicts how its actions will change the scene, imagining the next few moments of video and the actions that go with them in one diffusion transformer.
It succeeds the GR00T N1 series, and one pretrained checkpoint covers parallel-gripper arms, dexterous hands and a humanoid; a faster Turbo variant predicts in 124 milliseconds on NVIDIA’s Jetson Thor robot computer.
The benchmark numbers — and what they actually show
On 50 tasks the model had not been trained for, NVIDIA reports GR00T 2 completed 40 of them at least once with no task-specific demonstrations, averaging 40% success. With one demonstration per task, average success rose to 71%; with five — under three minutes of robot data per task, plus about 30 minutes of fine-tuning on a single RTX 5090 — it reached 95%.
Those figures need the usual caveat: they are NVIDIA’s own, from controlled tests, and the 95% figure carries a selection effect. NVIDIA’s whitepaper notes the five-demonstration tasks were pre-screened to ones the model could already learn at 80% success or better, so the headline number is high partly by construction. The 40% zero-shot average is the better measure of how far the model generalises on its own. More striking is cross-body transfer: 65% success on a filter-paper setup learned purely from human demonstrations, and 60% on a toy dump-truck assembly transferred from a handheld gripper to a YAM robotic arm.
The same week also brought Google’s Gemini Omni 1.2 Flash video model — and Tencent’s Hy4 computer-use model, which teaches software agents to see and act on screens.
Learning from human video
The training mix explains the approach: 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation. Adding the human video raised real-robot task progress from 50% to 70% across 72 tasks, NVIDIA says.
Of that human video, 100,000 hours are being released openly as the Human-1M-100K dataset — video, language annotations and head-and-hand motion data, under a CC BY 4.0 licence. Two tools are open too: a Video to Data workflow that turns human video into simulated robot demonstrations, and EgoHand, an open model that reconstructs 3D hand motion from video. The model itself is not fully open yet, though: NVIDIA’s research page lists code as “Available in December” and states no licence or release date for the model weights.
Isaac Lab 3.0 and the industrial trials
Alongside the model, NVIDIA introduced Isaac Lab 3.0, its open-source robot-learning framework, adding the Newton physics engine, which models deformable objects like cables and cloth, and multi-jointed robotic hands. In one dexterous-manipulation benchmark, NVIDIA reports Isaac Lab 3.0 with Newton achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX. Isaac Lab 3.0 itself has been generally available since 20 October 2026 — the CoRL news pairs it with GR00T 2 rather than launching it.
The model is already in trials, though none is a production rollout. Nexuni is testing GR00T 2 for robotic laundry handling, Foxconn is applying it to server-manufacturing automation, and Siemens is evaluating it across industrial use cases. MicroAGI is using it to swap Ethernet cables with a YAM robotic arm, after reinforcement-learning post-training.
Why this matters
GR00T 2 matters less for any single benchmark than for what it changes about how robots are taught. If skills can be learned from video of people, adding a new task stops costing weeks of teleoperation — it takes a handful of demonstrations, or none. That is the difference between a robot that does one job in one factory and one retrained for the next job like a worker.
The honest reading is narrower than the headline. The strongest numbers are vendor-reported, from controlled tests, with the best results on pre-screened tasks. The model is not downloadable yet, and no production deployment has been announced. But the dataset is genuinely open, the human-video approach is backed by a large measured gain, and the trials involve real factory operators. The test now is the December code release and whether outsiders can reproduce NVIDIA’s numbers.
