Nvidia GR00T 2 robot model reaches 95% after five demonstrations

Nvidia GR00T 2 robot model reaches 95% after five demonstrations

11 November 2026: Nvidia's GEAR research group has introduced GR00T 2, a robot foundation model that learns tasks from human video and, after five demonstrations of a new task and a 30-minute fine-tune on a single RTX 5090, succeeds 95% of the time on 50 tasks it had not seen on the target robot. Without any demonstration the same model succeeds 40% of the time. Nvidia says code will be available in December.

Not yet independently verified. All figures are results Nvidia reports from its own lab; we have seen no independent test, and the page gives no price, licence or release terms beyond code being available in December. The page is headed November 2026 without a day. We will update this when it can be confirmed, and remove this note.

Predicting video and actions together

GR00T 2 belongs to what Nvidia calls world-action models. A single diffusion transformer denoises future video and a chunk of robot actions together: it reads recent video, the instruction and the robot's state, executes the chunk, and repeats. Small per-robot heads turn the plan into motor commands for each body, and the models run on parallel grippers, dexterous hands and humanoids.

Pretraining mixes three kinds of data: 402,730 hours of egocentric video of people at work, 10,184 hours of demonstrations with a handheld UMI gripper across 2,897 tasks, and 23,500 hours of robot teleoperation. Nvidia says 100,000 hours of the human video are openly available as Human-1M-100K on Hugging Face. In its test of scaling, task progress on 72 downstream tasks rose from 50% to 70% on the real robot as the share of human video grew from none to all of it.

Skills shown to humans, performed by robots

Nvidia calls the central result X-Adaptability: a task demonstrated only by humans, or with a handheld or wearable device, carries over to a robot that never practised it. A robot with two five-fingered hands lined a lab funnel with filter paper and wet it with a wash bottle in 65% of trials after seeing only people do it. A YAM arm assembled a toy dump truck end to end in 60% of trials after seeing only a handheld UMI gripper, a task that takes up to four minutes.

Where it works and where it falls short

With a small amount of post-training, two 22-DoF Sharpa hands seated GB300 memory modules in 85% of trials and assembled a toy truck in 65%. Reinforcement learning on the real robot lifted three contact-rich insertion tasks to 88 to 96%, from 40 to 76% after supervised training.

When a reasoning model called GR00T 2 as a tool, the YAM arm completed 98, 97 and 97 of 100 assemblies of a toy plane, tractor and car, against 71 to 84 without the agent. Nvidia also reports limits: on a second robot, Dexter, marks that steer the model place more items in trash sorting but rarely complete the full task, which the page calls the current limit of the system.

A smaller model for the humanoid itself

GR00T 2 Turbo uses a 2-billion-parameter transformer and a more compact video encoder, and runs on the Unitree G1 humanoid's own Jetson Thor. A prediction takes 124 ms there and 38 ms on an RTX 5090. On the G1 it served popcorn in all 48 trials and cleared a table of diverse objects in 93% of trials. Adding simulated data raised its out-of-distribution success from 6.7% to 63.3%.

Nvidia reports that results in its simulated twins tracked real-world results closely, with a Pearson correlation of 0.92 on matched-task success across 14 policies on a 25-task YAM benchmark.

For anyone planning a robot cell, the numbers show what a few minutes of demonstration data might buy, but they come from one lab's chosen tasks. The release terms, which decide what a buyer can do with the model, have not been published.

Sources

• NVIDIA Research GEAR: GR00T 2, Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation (read 11 November 2026), 11 November 2026