Nvidia GR00T 2 is an open robot foundation model that predicts what a robot will see next and which actions to take, both in one pass. Nvidia unveiled it on November 9, 2026 at the Conference on Robot Learning (CoRL) in Austin, Texas, alongside Isaac Lab 3.0. The big claim: a robot can learn a task by watching people do it, then perform that task without ever practicing it.
Here’s how GR00T 2 works, what Nvidia’s numbers actually say, and where it still falls short.
What Is Nvidia GR00T 2?
Nvidia GR00T 2 is a world action model (WAM) for robots. It jointly predicts future video and robot actions, so it learns how physical interactions unfold along with the movements that cause them. It runs on parallel grippers, five-fingered dexterous hands, and humanoid robots, and it’s built on Nvidia’s Cosmos world models.
GR00T 2 at a glance
| Detail | What Nvidia says |
| Announced | November 9, 2026, CoRL, Austin, Texas |
| Model type | World action model (one diffusion transformer) |
| Built on | Nvidia Cosmos |
| Predecessor | GR00T N1 series (latest: N1.7) |
| Pretraining data | 402,730 hours of human video, 10,184 hours of handheld-gripper demos, 23,500 hours of robot teleoperation |
| Robot bodies | Parallel-gripper arms, dexterous hands, Unitree G1 humanoid |
| Code | Listed as available in December 2026 |
How Does GR00T 2 Work?
GR00T 2 takes recent video, a text instruction, and the robot’s current state. It then denoises a short stretch of future video and a chunk of robot actions together, in a single diffusion transformer. The robot executes that action chunk, and the loop starts again.
The video prediction isn’t decoration. Predicting how a scene should change is how the model picks up physics from human footage. Small per-robot output heads then turn the plan into motor commands for each specific body.
If you’ve followed earlier GR00T releases, this is a real shift. GR00T N1.7 is a vision-language-action (VLA) model: it looks at images and language, then outputs actions directly. A world action model adds an imagined future in the middle. Nvidia first previewed this approach at GTC in March 2026 under the name GR00T N2, based on its DreamZero research. The shipped name is GR00T 2.
GR00T 2 vs GR00T N1.7
| GR00T N1.7 | GR00T 2 | |
| Approach | Vision-language-action | World action model |
| What it predicts | Actions | Future video and actions together |
| Backbone | Cosmos-Reason2-2B vision-language model plus diffusion action head | Joint video-action diffusion transformer on Cosmos |
| Learns from human video | Limited | Core of the training mix |
What Is a World Action Model?
A world action model is an AI model for robotics that learns both how the world is likely to change and what actions a robot can take to shape that change. Instead of mapping a camera image straight to a motion, it predicts the outcome it is acting toward, which helps it plan for new tasks and unfamiliar scenes.
What Data Was GR00T 2 Trained On?
Three sources, in very uneven amounts.
The bulk is 402,730 hours of first-person video of people at work, drawn from Nvidia’s Human-1M dataset. Next comes 10,184 hours of demonstrations from a handheld UMI gripper across 2,897 tasks. The smallest slice is 23,500 hours of robot teleoperation from arms, dexterous hands, and the Unitree G1 humanoid.
Nvidia ran a scaling test on this. It pretrained four models on none, 1%, 10%, and 100% of the human video, then tested them on 72 tasks. Real-robot task progress rose from 50% to 70% as human video went up. That’s the strongest argument in the whitepaper for why watching people matters.
A 100,000-hour portion of the human video is public as Human-1M-100K on Hugging Face.
What Can GR00T 2 Actually Do?
It learns from humans, then runs on robots that never practiced the task
Nvidia calls this cross-adaptation, or X-Adapt. It shows three versions:
- Human to robot: Shown only by gloved people with head and wrist cameras, a robot with two five-fingered hands lined a lab funnel with filter paper and wet it in 65% of trials.
- Handheld gripper to robot arm: Shown only with a handheld UMI gripper, a YAM arm assembled a toy dump truck end to end in 60% of trials. That task takes up to 4 minutes.
- Wearable hand to robot hand: Using a DexUMI exoskeleton, a two-handed robot made banana yogurt, scooped granola, and struck a match to light a lamp.
Zero-shot and few-shot results
Nvidia tested 50 tasks that appeared only in human video or other robots’ data, never on the target robot (a bimanual robot with two Sharpa hands).
| Setting | Average success across 50 tasks |
| Zero-shot (instruction only) | 40% |
| 1 demonstration per task | 71% |
| 5 demonstrations per task | 95% |
Five demonstrations is under 3 minutes of robot data per task. Each adaptation was a 30-minute LoRA fine-tune on a single RTX 5090.
Dexterous, contact-heavy work
With small amounts of post-training data, GR00T 2 handled long tasks with two 22-DoF Sharpa hands: seating GB300 memory modules (85% success) and assembling a toy truck (65%). Nvidia says earlier models couldn’t do these. Reinforcement learning on the real robot, using only task-success signals, lifted three insertion tasks to 88-96%.
Agentic control
A reasoning model can call GR00T 2 through a WAM API, the way an agent calls a tool. The agent decides what to do. GR00T 2 handles how. With that setup, the YAM arm assembled a toy plane, a tractor, and a Hot Wheels car in 98, 97, and 97 of 100 trials. The same policy without the agent managed 71-84.
Visual steering goes a step further: the agent marks grasp and release points on GR00T 2’s camera view. That raised success on “what and where” instructions from 40% to 92% on YAM.
What Is GR00T 2 Turbo?
GR00T 2 Turbo is a smaller, faster version of the model, built to run on the robot itself. It uses a 2-billion-parameter transformer and a more compressed video representation. On a Unitree G1 humanoid, one prediction takes 124 milliseconds on the onboard Jetson Thor. On an RTX 5090 it takes 38 ms.
It steers the SONIC whole-body controller, so the robot can walk, bend, reach, and carry at once. In Nvidia’s tests, the G1 scooped and served popcorn in all 48 trials and cleared a table of mixed objects in 93% of trials.
Simulation helped a lot here. Adding simulated episodes pushed Turbo’s out-of-distribution success on the G1 from 6.7% to 63.3%. GR00T 1.7 only reached 40.0% with the same extra data.
How Does GR00T 2 Rank?
Nvidia says GR00T 2 ranks among the top three models across all vision-language-action and world action model categories on RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces, with several leaderboard wins. Those are Nvidia’s claims. Treat them as a starting point until outside teams reproduce them.
Is GR00T 2 Open Source?
Nvidia calls GR00T 2 an “open frontier world action model,” and the research page lists code as available in December. Part of the training data, Human-1M-100K, is already public. The exact license terms and weight downloads weren’t spelled out in the material I found, so check Nvidia’s GR00T page and Hugging Face collection when the code lands.
What Are GR00T 2’s Limits?
The whitepaper is fairly upfront about these, which I appreciate.
- Zero-shot is 40%: That’s impressive for unseen tasks and still a failure rate of 60%. Most misses come from precision, not from the wrong plan.
- Long tasks on dexterous hands are unsolved: With visual steering, the Dexter robot placed more items in trash sorting but rarely finished the full task.
- Marks don’t carry everything: They don’t convey an object’s final orientation or motions like pouring and wiping. Those still come from language.
- Turbo trades accuracy for speed: It trails the full model on four of six YAM tasks and five of six Dexter tasks.
- The benchmarks are Nvidia’s own: Robot labs, tasks, and hardware are chosen by the team that built the model.
So is this a robot “brain”? Sort of. It’s closer to a skilled hand-and-eye system that a smarter planner can direct. The reasoning model picks the task, GR00T 2 imagines and plans the motion, and a controller moves the joints.
Why GR00T 2 Matters for Robotics
Collecting robot teleoperation data is slow and expensive. Human video is cheap by comparison, and there’s an enormous amount of it. If a model can turn footage of people working into skills for a robot it has never seen, the cost of teaching a new task drops. A 30-minute fine-tune on one consumer-grade GPU is the kind of detail that makes small labs pay attention.
Whether it holds up outside Nvidia’s test setups is the open question. The December code release will be the first real chance to find out.
GR00T 2 FAQ
What is Nvidia GR00T 2?
Nvidia GR00T 2 is an open world action model for robots, announced November 9, 2026. It predicts future video and robot actions together in one diffusion transformer, which lets it learn skills from human video and transfer them to grippers, dexterous hands, and humanoids.
When was GR00T 2 released?
Nvidia introduced GR00T 2 on November 9, 2026 at the Conference on Robot Learning in Austin, Texas. The research page lists code as available in December 2026. Nvidia previewed the same architecture at GTC in March 2026 under the name GR00T N2.
What is the difference between GR00T 2 and GR00T N1.7?
GR00T N1.7 is a vision-language-action model that outputs actions directly from images and language. GR00T 2 is a world action model that predicts future video and actions together. That lets it learn physical behavior from human video and adapt to new robots with far less robot data.
Can GR00T 2 run on a humanoid robot?
Yes, in its Turbo version. GR00T 2 Turbo runs onboard a Unitree G1 humanoid using a Jetson Thor, with each prediction taking 124 milliseconds. Nvidia says the robot served popcorn in all 48 trials and cleared a table of diverse objects in 93% of trials.
Does GR00T 2 learn from human videos?
Yes. Its pretraining mix includes 402,730 hours of first-person human video, the largest share of the data. Nvidia reports that robots can perform tasks demonstrated only by humans, such as a lab funnel setup, in 65% of trials without any teleoperated demonstrations.
What hardware do you need to fine-tune GR00T 2?
Nvidia reports adapting GR00T 2 to a new task with a 30-minute LoRA fine-tune on a single RTX 5090, using five demonstrations. Running Turbo on a humanoid used a Jetson Thor onboard. Full requirements will be clearer once the code is released in December.
