Nvidia introduces GR00T 2, a robot AI trained on human video

Robots can learn manipulation skills from human footage and improve with just a few task demonstrations, according to Nvidia’s results for GR00T 2. The company introduced the model alongside Isaac Lab 3.0 at the Conference on Robot Learning in Austin on November 9, 2026, pairing a new robot model with tools for practicing and refining skills in simulation.

Predicting the scene and the movement

GR00T 2 does two connected jobs: predicting what a robot’s camera will see next and generating the movements that accompany that view. Nvidia calls this a world action model. Built on its Cosmos world models, it takes recent video, an instruction and the robot’s current state, then produces short sequences of future video and motor commands together.

The predictions share one diffusion transformer, the model architecture that generates both outputs, with small output components tailored to each robot. A faster variant, GR00T 2 Turbo, uses a two-billion-parameter transformer. Nvidia reports that it produces a prediction in 124 milliseconds on the company’s Jetson Thor robot computer.

Its training mix extends well beyond robot recordings. Nvidia lists 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations and 23,500 hours of remotely operated robot activity. The robot data spans arms, dexterous hands and Unitree’s G1 humanoid. Nvidia also released 100,000 hours of human footage as an open dataset on Hugging Face.

A first success is not a success rate

Nvidia reports that adding all the human video raised real-robot task progress from 50% to 70% across 72 tasks. That is a separate measure from its evaluation on 50 unseen tasks, where it tested how performance changed as the model received more demonstrations.

Without task-specific demonstrations, GR00T 2 completed 40 of those 50 tasks at least once. That does not mean an 80% average success rate: Nvidia’s research results put the average at 40%. With five demonstrations per task, the company reports reaching its strongest result after a 30-minute fine-tune on a single RTX 5090, using under three minutes of robot data per task. These are Nvidia’s own results, not independent validation.

40%No demonstrations

Nvidia reports 40% average success without task-specific demonstrations.

71%One demonstration per task

Nvidia reports 71% average success with one demonstration per task.

95%Five demonstrations per task

Nvidia reports 95% average success with five demonstrations per task.

Practicing cables, cloth and robot hands

Isaac Lab 3.0 supplies the simulation side of the announcement. Researchers can use AI agents to set up training environments, switch between the PhysX and Newton physics engines, and apply a new reinforcement-learning library to refine models such as GR00T through practice. Newton integration supports deformable objects, including cables and cloth, and complex robot hands with many interacting joints.

Nvidia says Isaac Lab 3.0 with Newton delivered up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX on a vision-based dexterous-manipulation benchmark. The comparison changes both the framework version and physics engine. For evaluation, the Isaac Lab-Arena 0.4 developer preview lets teams run robot policies—the software controlling their behavior—across thousands of variations in parallel to identify failure conditions.

Early applications named by Nvidia

• Foxconn is applying GR00T 2 to factory automation for server manufacturing.
• Siemens is evaluating the model for industrial use cases.
• Nexuni is testing it for robotic laundry handling.

Access remains a material qualification. Nvidia describes GR00T 2 as open, but its research page lists code as available in December without specifying a model-weight release date or license, according to MLQ. The released human-video dataset and the introduced robot model therefore have different availability details; the announcement does not provide a complete timetable for developers to obtain and use the model itself.