NVIDIA has unveiled GR00T 2, a new open robot foundation model that learns dexterous hand skills by watching first-person video of people doing everyday tasks. The company introduced the model on Monday 9 November 2026 at the Conference on Robot Learning (CoRL) in Austin, Texas, alongside a major update to its robot-training software.

GR00T 2 is built on NVIDIA Cosmos, the company's platform for physical AI. It was trained on a mix of robot data and human data, including 100,000 hours of open-source first-person human video released by NVIDIA, so robots can pick up manipulation skills they have never practised on their own bodies.

Robots that learn by watching

NVIDIA describes GR00T 2 as a world action model: it predicts how its actions will change what it sees before it moves, rather than simply reacting to each frame. In the company's reported evaluation, the model completed 40 of 50 tasks at least once without any task-specific demonstrations. With one demonstration per task, it averaged 71 per cent success across the 50 tasks; with five demonstrations, that rose to 95 per cent. Those figures come from NVIDIA's own defined test conditions and have not been independently verified.

The model is also designed to transfer skills between different robot bodies without needing new robot demonstrations, and AI agents can use its programming interface to run skills in sequence and recover from errors. NVIDIA says GR00T 2 ranks among the top three models across the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins.

Simulation does the rehearsing

Alongside the model, NVIDIA introduced Isaac Lab 3.0, the latest version of its open-source robot-learning framework, which has been generally available since 20 October 2026. It lets researchers use AI agents to set up training environments quickly, switch between the PhysX and Newton physics solvers, and apply a new reinforcement post-training library to foundation models such as GR00T 2. In a vision-based dexterous manipulation benchmark, NVIDIA says Isaac Lab 3.0 with Newton reached up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX.

Accurate physics matters because skills learned in simulation must still work on a real robot. The tighter Newton integration lets Isaac Lab model deformable objects such as cables and cloth, as well as robotic hands with many interacting joints. NVIDIA also released two new open benchmarks, DexBench and VABAR, for measuring robot dexterity on industrial tasks, and a developer preview of Isaac Lab-Arena 0.4 for scaled policy evaluation. A new open Video to Data workflow turns human video into simulated robot demonstrations.

Early trials, not yet factory floors

The first real-world tests are already under way, but none is a production rollout. Nexuni is testing GR00T 2 for robotic laundry handling, Foxconn is applying it to factory automation for server manufacturing, and Siemens is evaluating it for industrial use. MicroAGI is using the model, post-trained through reinforcement learning, to perform Ethernet cable swapping with a YAM robotic arm.

Ken Goldberg, professor of robotics at UC Berkeley and a lead on the VABAR and Graph-as-Policy projects, said: "To me, agentic robotics is the most exciting technological development since the internet in 1995."

The announcement gives no pricing or general availability date for GR00T 2. The open question now is whether benchmark results achieved in simulation and controlled tests translate into reliable performance on factory floors, where a task that fails even one time in twenty can still disrupt a production line.

Sources: NVIDIA Blog — Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation and The Robot Report — NVIDIA launches GR00T 2 and Isaac Lab 3.0 at CoRL.