Nvidia Unveils GR00T 2 and Isaac Lab 3.0 to Teach Robots From Human Video

Published November 10, 2026 | Tech & AI | TrendWire Daily

Nvidia wants robots to learn the way people do — by watching. On Monday, November 9, 2026, the company introduced GR00T 2, its updated open robot foundation model, and Isaac Lab 3.0, the latest version of its robot-learning framework, at the Conference on Robot Learning (CoRL) in Austin, Texas.

What Nvidia Announced at CoRL

Direct answer: Nvidia unveiled GR00T 2, an open frontier world action model for humanoid robots, and Isaac Lab 3.0, its open-source robot simulation and training framework, at CoRL 2026 in Austin on November 9. GR00T 2 learns dexterous manipulation from human video: Nvidia reports it completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations, and reached 95 percent average success across 50 tasks with just five demonstrations each. Isaac Lab 3.0 adds faster simulation, switchable physics engines and a new reinforcement post-training library.

CoRL 2026 runs from November 9 to 12 in Austin. Jim Fan, co-lead of Project GR00T, gave a keynote on November 9 titled "From Dreaming to Doing: World Action Models and Agentic Robotics." The announcements position simulation — robots rehearsing delicate work in virtual worlds before touching real hardware — as the core of Nvidia's robotics strategy.

GR00T 2: Learning Dexterity From Human Video

GR00T 2 is built on Nvidia Cosmos, the company's world-model platform, and uses a 2-billion-parameter transformer. It was trained on diverse robot and human data, including 100,000 hours of Nvidia open-source egocentric (first-person) human video, plus thousands of hours of handheld-gripper demonstrations and robot teleoperation. The idea is that a robot that has watched humans do thousands of tasks can generalize to new ones with only a handful of examples.

The numbers Nvidia reported are striking — with an important caveat. On 50 evaluated manipulation tasks, GR00T 2 completed 40 at least once with zero task-specific demonstrations (a 40 percent zero-shot success rate). With one demonstration per task, average success rose to 71 percent, and with five demonstrations it reached 95 percent. Nvidia also says GR00T 2 ranks among the top three models across vision-language-action and world-action-model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins.

The caveat, noted in Nvidia's technical whitepaper, is that the five-demonstration tasks were selected among those the model can already learn with five episodes at a success rate of at least 80 percent — so that average is high partly by construction. These are Nvidia's own claims, not independent results, and the code will not be available until December, according to Nvidia's research page.

Isaac Lab 3.0: The Simulation Training Ground

Isaac Lab 3.0 is the training ground where GR00T 2's policies are refined. The framework became generally available on October 20, 2026, a few weeks before CoRL, and the new release adds a modular architecture: researchers can switch between Nvidia's PhysX and Newton physics solvers, choose renderers for perception-based training, and use agents to quickly set up training environments with simulation-ready assets.

In a vision-based 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. Newton also models deformable objects such as cables and cloth — the kind of soft, fiddly materials that have long defeated robots. The release includes a new reinforcement post-training library for foundation models such as GR00T, plus deployment tooling called LEAPP to move a trained policy from simulator to real hardware.

Nvidia also introduced Isaac Lab-Arena, an evaluation layer with a new open benchmark for testing robot policies in simulation, and the developer preview of version 0.4 arrived at CoRL. Two Nvidia papers are among six finalists for the Best Paper Awards at CoRL, out of thousands of submissions.

Real-World Tests: From Laundry to Server Factories

The demos at CoRL focused on precision work: aligning parts, connecting cables, assembling small components. MicroAGI is using GR00T 2, post-trained through reinforcement learning, to perform Ethernet cable swapping with a YAM robotic arm — a task that combines vision, touch and careful force control. Nexuni is testing GR00T 2 for laundry handling, Foxconn is applying it to server-manufacturing automation, and Siemens is evaluating it for industrial use cases.

Other partners named by Nvidia include Ambi Robotics, Bosch Research, Dexmate, Dyna Robotics, FieldAI, Flexion, Noitom Robotics, Skild AI, RLWRLD, UC Berkeley, Walden Robotics and the University of Washington. The disclosed work is testing and evaluation; no production rollout has been announced. Using a technique called Success-Guided Sampling, developed with the University of Washington, Nvidia says its AlphaNIST policy learned all 20 tasks on the NIST Assembly Task Board with more than 98 percent training success using roughly 39 times less GPU time than separate per-task policies.

Humanoid robot training in a simulated warehouse, handling cables and small parts

Why It Matters — and What Comes Next

The shift from locomotion to manipulation is the theme of this year's CoRL. Walking robots are increasingly solved; robots that can reliably pick up, assemble and handle objects are not. If GR00T 2's generalization claims hold up in independent testing, it could cut the biggest cost in robotics: the thousands of hours of demonstrations currently needed to teach each new task.

That matters commercially for Nvidia too. The company is expanding from chips into the full robotics stack — models, simulation, benchmarks and deployment — and it is doing so while under unusual pressure, with a Justice Department probe into alleged chip smuggling and an independent board investigation into CEO Jensen Huang both in the news. The robotics announcements show Nvidia pushing ahead with its product roadmap regardless. For a broader look at Nvidia's AI partnerships, see our report on Nvidia's sovereign-AI deal with Rebellions and SK Telecom in Korea.

Frequently Asked Questions

What is Nvidia GR00T 2?

GR00T 2 is Nvidia's updated open frontier world action model for humanoid robots, unveiled on November 9, 2026 at CoRL in Austin. Built on Nvidia Cosmos with a 2-billion-parameter transformer, it learns dexterous manipulation skills from robot and human video data.

What is the 50-task claim about GR00T 2?

Nvidia reports GR00T 2 completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations, reached 71 percent average success with one demonstration per task, and 95 percent with five demonstrations. The five-demonstration tasks were pre-selected for learnability, so the 95 percent figure is high partly by construction.

What is Isaac Lab 3.0?

Isaac Lab 3.0 is Nvidia's open-source robot-learning framework for simulation training. It became generally available on October 20, 2026. It lets researchers switch between PhysX and Newton physics solvers, train thousands of simulated robots in parallel, and apply reinforcement post-training to foundation models like GR00T.

How much faster is Isaac Lab 3.0?

Nvidia reports up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX in a vision-based dexterous manipulation benchmark, when Isaac Lab 3.0 runs with the Newton physics engine. This is Nvidia's own benchmark result.

When can developers use GR00T 2?

Nvidia describes GR00T 2 as open, but its research page says code will be available in December 2026. No exact release date has been announced.

Who is testing GR00T 2?

Nexuni is testing it for laundry handling, Foxconn is applying it to server-manufacturing automation, Siemens is evaluating industrial use cases, and MicroAGI is using it for Ethernet cable swapping. All disclosed work is testing and evaluation, with no production rollout announced.

Search This Blog

TrendWire Daily