NVIDIA's GR00T 2 Teaches Robots New Skills Just by Watching Humans

 

NVIDIA just gave robots a shortcut to dexterity: stop programming every movement, and let them watch us instead. On Monday, November 9, 2026, at the Conference on Robot Learning (CoRL) in Austin, Texas, NVIDIA unveiled GR00T 2 — a robot foundation model that learns new manipulation skills largely from videos of humans doing the job. The headline number says it all: trained partly on more than 400,000 hours of first-person human video, it completed 40 out of 50 brand-new tasks at least once with zero practice runs.

What GR00T 2 Actually Is

GR00T 2 — pronounced "groot two" — is what NVIDIA calls an open frontier world action model, built on its Cosmos world-model platform. It uses a 2-billion-parameter transformer that predicts what the robot's camera will see next and the actions to take, together, in a single step. Give it a recent camera view, a written instruction and the robot's current state, and it generates a short chunk of future video plus the motor commands to match. One pretrained checkpoint covers parallel-gripper arms, dexterous hands and a humanoid robot.

The Big Idea: Human Video Is the Training Data

Robot learning has always had a data problem. Teaching a robot by demonstration means expensive teleoperation sessions, one task at a time. Human video, by contrast, is everywhere. GR00T 2 flips the balance: its pretraining mix includes 402,730 hours of first-person human video — people doing everyday tasks filmed from a head-mounted camera — alongside 10,184 hours of handheld gripper demonstrations across 2,897 tasks and 23,500 hours of robot teleoperation. NVIDIA is also releasing 100,000 hours of that human video as an open dataset called Human-1M-100K, under a Creative Commons licence that allows commercial use. Its own study found the human video mattered: adding all of it lifted real-robot 

task progress from 50% to 70% across 72 tasks.

The Numbers, With One Honest Caveat

On 50 tasks the robot had never seen, NVIDIA reports GR00T 2 averaged 40% success with no demonstrations — completing 40 of them at least once — 71% with a single demonstration per task, and 95% with five. It also has a trick NVIDIA calls X-Adapt: learn a skill from human video, or from a completely different robot, then perform it on a target robot that has never done the task. The caveat worth stating plainly: these are NVIDIA's own test numbers, and the five-demo figure comes from tasks pre-selected as learnable. Also, NVIDIA calls GR00T 2 open, but the code arrives in December and no public release date or licence for the model weights has been announced yet — open, with a short waiting room.

The Ecosystem Around It

GR00T 2 did not arrive alone. NVIDIA also highlighted Isaac Lab 3.0, its open robot-learning framework, plus two new open benchmarks — DexBench and VABAR — for testing robots on industrial tasks, and a Video to Data workflow that converts human video into simulated robot demonstrations. Early testers named by NVIDIA include Nexuni, trying it for laundry handling, Foxconn, for server-manufacturing automation, and Siemens, for industrial use cases. No production rollout has been announced yet — this is testing and evaluation.

Why It Matters

If GR00T 2's approach holds up outside NVIDIA's lab, the bottleneck for useful robots shifts. Instead of hand-collecting demonstrations for every chore, robot makers could tap the vast library of how humans already do things — the same way language models tapped the internet's text. The timing fits the week: Anthropic's Claude Fable 5.2 and Meta's Muse Ember both landed on November 9 too, and the model race is increasingly about what AI can do in the physical world, not just on a screen. Robots that learn by watching us are still early, and the numbers are the vendor's own — but the direction is clear. The next robot you meet may have learned its job the same way you did: by watching someone else do it first.