Nvidia's GR00T 2 robot model predicts future video and actions together, and learns mostly from human footage
Nvidia presented GR00T 2 at the Conference on Robot Learning on 9 November. Its research page reports that more human video improved robot performance, and that a handful of demonstrations lifted success on unseen tasks.

Nvidia presented GR00T 2, a new family of robot foundation models from its GEAR research lab, at the Conference on Robot Learning (CoRL) in Austin, Texas, on 9 November 2026. The company describes it as a “world action model”: rather than mapping what a robot sees straight to motor commands, a single network predicts what the scene will look like next and what the robot should do, at the same time. Most of its pretraining data is video of people doing physical work, not robot data.
The research page and an accompanying whitepaper set out the design and a series of experiments. All results below are Nvidia’s own and have not been independently replicated.
Predicting the world, not just the action
Many recent robot models are vision-language-action (VLA) models: they take camera images and an instruction and output actions. GR00T 2 adds a second job. One diffusion transformer denoises a short clip of future video together with a chunk of robot actions, and small robot-specific components turn that plan into motor commands. The robot executes the chunk, looks again, and repeats. Nvidia’s blog post says the model is built on Nvidia Cosmos, the company’s family of world foundation models.
The research team frames the bet in one line: “Learning to predict the world, not just act in it, unlocks three new capabilities.” The page lists them as cross-embodiment adaptation (“X-Adaptability”), fast adaptation with agentic control, and whole-body loco-manipulation using a smaller variant.
The data mix
According to the research page, pretraining draws on three sources:
| Data source | Hours |
|---|---|
| Egocentric video of people at work | 402,730 |
| Demonstrations with a handheld UMI gripper (2,897 tasks) | 10,184 |
| Robot teleoperation | 23,500 |
“Human video is by far the largest part of the mixture,” the page notes. Nvidia has released a 100,000-hour slice of that footage as Human-1M-100K on Hugging Face under a CC BY 4.0 licence. The dataset card lists about 3.4 million episodes with language annotations, head-camera pose and 21 hand keypoints per hand.
Does more human video help?
This is the core scientific question. To test it, the team pretrained four models on none, 1%, 10% and 100% of the human video and evaluated them on 72 tasks on a robot it calls Dexter, fitted with Sharpa hands. Nvidia reports that task progress (a partial-credit measure, not full success) “rises at every step, from 64% to 77% in video rollouts and from 50% to 70% on the robot”.
Unseen tasks and fast adaptation
Nvidia also tested 50 tasks that appear in other robots’ data or in human video “but never in the target robot’s own data”, using a bimanual robot with two Sharpa hands:
- Zero-shot, from the instruction alone, the model succeeded in 40% of trials; Nvidia’s blog adds that it completed 40 of the 50 tasks at least once.
- One demonstration per task raised average success to 71%.
- Five demonstrations per task, under three minutes of robot data, raised it to 95%.
Each adaptation was a 30-minute LoRA fine-tune on a single RTX 5090 graphics card. In a separate cross-embodiment test, the page says that learning from human demonstrations alone, with no teleoperation, “GR00T 2 succeeds on all four tasks”. A smaller variant, GR00T 2 Turbo, uses a 2-billion-parameter transformer and is described as light enough to run on board a Unitree G1 humanoid.
Stated limitations
The research page is candid about where the system falls short:
- On unseen tasks, “most misses come from execution, not intent”: the model picks the right steps in the right order, but a single step lacks precision.
- Visual marks used to steer the robot convey neither an object’s final orientation nor the motion between grasp and release, so language is still needed for tasks such as pouring or wiping.
- On Dexter, visual marks helped the robot place more items in a trash-sorting task but rarely completed the full task, which the page calls “the current limit of the system”.
- GR00T 2 Turbo trails the full model on most of the tasks reported, including five of six on Dexter, though it comes out ahead on two tasks on a YAM arm.
Availability
The whitepaper is available now, and the research page says code will be “Available in December”. The Human-1M-100K dataset is already downloadable. Nvidia’s blog describes GR00T 2 as open and says Foxconn is applying it to server manufacturing automation and Siemens is evaluating it for industrial use; neither page gives a date for model weights.
What it means
Robot learning is short of data. Teleoperating a robot to record demonstrations is slow and costly, while video of people handling objects is comparatively plentiful. Human footage, though, contains no robot motor commands, so the question has been how much a robot can actually learn from it. Training a model to predict future frames is one answer: it can learn how objects move and respond to hands from video alone, then map that understanding onto a particular robot with a small amount of robot data.
Nvidia’s scaling experiment is the most interesting evidence here, because it isolates the variable: same tasks, same robot, more human video, better results. The caveats are the usual ones for a vendor whitepaper. The numbers come from Nvidia’s own robots and task sets; the 95% figure follows task-specific demonstrations, while the zero-shot figure is 40%; and outside groups cannot reproduce the training pipeline until the code arrives. Independent evaluation on other hardware will show how far the human-video approach generalises.
Sources
- Nvidia GEAR: GR00T 2 research page
- Nvidia Blog: Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation (9 November 2026)
- Hugging Face: nvidia/PhysicalAI-Human-1M-100K dataset card
- The Robot Report: coverage of GR00T 2 and Isaac Lab 3.0 at CoRL (9 November 2026)



