NVIDIA launches GR00T 2, trained on 100,000 hours of human video

Key takeaways

• NVIDIA introduced GR00T 2, an open world action model built on NVIDIA Cosmos, at the Conference on Robot Learning (CoRL) in Austin, Texas.
• The model was trained on robot data plus 100,000 hours of first-person human video, which NVIDIA is releasing as open data.
• Isaac Lab 3.0 adds agent-assisted setup, Newton physics for cables and cloth, and a reinforcement learning library for post-training.

NVIDIA has launched GR00T 2, a robot foundation model that learns manipulation skills from human video and moves them between robots with different bodies. The company announced it on 9 November 2026.

The launch took place at CoRL in Austin, Texas, alongside Isaac Lab 3.0, the newest version of NVIDIA's open-source robot learning framework. NVIDIA presented both as tools to help robots move from lab demonstrations to factory work.

What GR00T 2 is

NVIDIA describes GR00T 2 as a "world action model". Before a robot moves, the model forecasts how that movement will change the scene in front of the robot's cameras.

According to NVIDIA Research, one diffusion transformer generates future video and robot actions together. Small heads for each robot then turn that plan into motor commands for a gripper arm, a dexterous hand or a humanoid.

The robot foundation model sits on NVIDIA Cosmos, the company's family of world models. Its training mix combines robot data with 100,000 hours of egocentric footage, meaning video filmed from a person's point of view as they handle objects. NVIDIA has published that footage on Hugging Face as an open dataset.

How the robot foundation model performed in NVIDIA's tests

NVIDIA reported results across 50 tasks it selected for evaluation. With no task-specific demonstrations, GR00T 2 completed 40 of those tasks at least once.

Given a single demonstration per task, average success reached 71%. With 5 demonstrations, that rose to 95%. NVIDIA Research said those 5 demonstrations amount to less than 3 minutes of robot data, followed by a 30-minute fine-tune on one RTX 5090 graphics card.

AI agents can also call GR00T 2 through an application programming interface. NVIDIA said this lets an agent chain skills together and recover when a step fails. The company added that GR00T 2 ranks in the top three models on several public robot benchmarks, including RoboCasa365 and RoboLab.

These figures come from NVIDIA and have not yet been reproduced by outside groups.

Who is testing the robot foundation model

NVIDIA named three early users. Nexuni is testing the robot foundation model for robotic laundry handling. Foxconn is applying it to automation for server manufacturing, and Siemens is evaluating it across industrial uses.

MicroAGI is using a version post-trained with reinforcement learning to swap Ethernet cables with a YAM robot arm. None of these companies has published production results.

Isaac Lab 3.0 and simulation training

Isaac Lab is where developers train and check robot skills in software before a physical machine tries them. The 3.0 release lets AI agents build training environments from simulation-ready assets.

Users can now move between NVIDIA PhysX and Newton, an open-source physics engine. Newton can simulate soft and flexible objects such as cables and cloth, plus robot hands with many joints. A new reinforcement learning library supports post-training of models such as GR00T.

In one vision-based dexterous manipulation benchmark, NVIDIA said Isaac Lab 3.0 with Newton delivered up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX.

Partners are building on the same engine. Skild AI, co-founded by Deepak Pathak and Abhinav Gupta, introduced STEEL, a rigid-body physics solver built on Newton to speed up training in Isaac Lab.

Agentic robotics and the view from UC Berkeley

NVIDIA also showed Graph-as-Policy, a UC Berkeley framework where AI agents link individual skills into robot programs, test them in simulation and fix them after failures. Engineers can read those programs before deployment.

Ken Goldberg leads the project at UC Berkeley, where he holds the William S. Floyd Jr. Distinguished Chair of Engineering. He is also co-founder and chief scientist of Ambi Robotics and co-founder of Jacobi Robotics, and he is a member of the National Academy of Engineering.

Goldberg said, "To me, agentic robotics is the most exciting technological development since the internet in 1995."

NVIDIA's robotics work is led by Deepu Talla, vice president and general manager of Robotics and Edge AI. Rev Lebaredian is the company's vice president of Physical AI Simulation.

A growing market for factory robots

Demand for factory automation gives any robot foundation model a large potential base. The International Federation of Robotics (IFR) counted 542,000 industrial robots installed worldwide in 2024, more than double the figure 10 years earlier.

The IFR put the global operational stock at 4.66 million units in 2024, up 9% on the previous year. It expects annual installations to pass 700,000 by 2028.