Watching someone fold laundry and successfully folding laundry are two very different achievements. Anyone with a chair full of clean clothes understands this.
NVIDIA is trying to close that gap for robots.
On November 9 at the Conference on Robot Learning in Austin, the company introduced GR00T 2, a robot foundation model built on NVIDIA Cosmos. It learns from human video alongside handheld demonstrations and robot training data, then predicts future video and robot actions together.
The bigger bet: robots could learn much of what work looks like by watching people, reducing how much teaching has to happen on expensive physical machines.
That lands directly inside robotics’ most consequential argument: Will useful robots arrive by collecting vastly more demonstrations, or by finding better ways to program them?
At CoRL, NVIDIA is backing both approaches.
The training mix tells the story
NVIDIA calls GR00T 2 a world action model. Its architecture uses one diffusion transformer to jointly predict how a scene will change and which robot actions will produce that change.
In plain English: it learns to connect a movement with what should happen next. Imagine reaching for a cup. The useful lesson includes how the cup should move when the grasp works.
According to NVIDIA’s research page, the training mixture includes:
- 402,730 hours of first-person human video.
- 10,184 hours of demonstrations using a handheld gripper.
- 23,500 hours of robot teleoperation, where people remotely control robots.
The human footage dominates. But the robot and gripper demonstrations remain part of the recipe.
That distinction matters. This is evidence for combining human observation with robot experience, rather than proof that robots can learn everything they need from watching us.
The attraction is easy to understand. If filming people doing useful work can replace some robot-specific data collection, teaching robots could become easier to scale. You could capture variation across workplaces without sending a robot to every location.
We’ve covered a related strategy in Sunday Robotics’ approach to collecting human demonstrations for Memo. Different systems, similar economic question: how much useful training can happen before a robot enters the room?
Read the 95% figure carefully
NVIDIA reports that GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations on the target robot.
For clarity, that's task coverage, not an 80% success rate.
The research page separately reports a 40% zero-shot trial success rate. With one demonstration per task, average success rose to 71%. With five demonstrations, it reached 95%, after a 30-minute fine-tune on a single RTX 5090. The tasks could already appear in human video or data from other robots; they were unseen on the target robot. NVIDIA describes the evaluation here.
Those are promising adaptation results. They also come with boundaries welded on.
Every performance figure above is NVIDIA’s own. The announcement does not name the 50 tasks or disclose the rollout counts for that evaluation.
Without those details, readers cannot judge the breadth of the task selection or how much evidence sits behind each average.
A factory operator also needs answers that a pooled success score cannot provide. What happens when a part slips? Does the robot recover? How often does someone intervene? Can it repeat the job throughout a shift?
The meaningful claim is that a pretrained model may need relatively little additional robot data to adapt. Whether that produces dependable, economical work remains a separate test.
Meanwhile, another camp wants robots to write code
The timing makes this launch more interesting.
In a November 9 IEEE Spectrum feature, Evan Ackerman examines an alternative approach from UC Berkeley and NVIDIA: agents that write robot-control programs, test them in simulation, inspect failures, and revise the code.
That approach shifts some of the burden from collecting demonstrations to generating and improving programs using existing robotics tools. It also produces code engineers can inspect.
The constraint moves with it. A program refined in simulation still has to survive physical reality, including contact and friction that the simulator may represent imperfectly.
These approaches ask different things of scale. GR00T 2 asks how much human experience a model can absorb and transfer. Agentic robotics asks how much programming and testing an AI system can automate.
Our read: the practical opportunity is in combining them.
NVIDIA’s own GR00T 2 research already describes a reasoning agent that sequences skills, checks progress, and invokes recovery tools. A learned model handles manipulation while the agent coordinates the work. The research page explains that division of labor.
For buyers, the useful question becomes which combination can perform a particular job reliably—and how much effort it takes to get there.
“Open” still needs a release date and terms
There is an accessible piece today: NVIDIA’s Human-1M-100K dataset, containing 100,000 hours of human video under CC BY 4.0.
That is a subset of the 402,730 human-video hours used in GR00T 2’s training mixture.
The model’s research page lists code availability for December 2026. It does not establish a weights release date or publish the model’s license terms. As of November 10, this should not be described as a released open-weight model.
That distinction affects what outside teams can reproduce, adapt, and deploy. As we explored in NVIDIA and Hugging Face’s push toward shared robotics infrastructure, broader participation depends on usable releases and clear permissions.
The surrounding CoRL announcements include Isaac Lab 3.0, the DexBench and VABAR benchmarks, and evaluations involving Foxconn, Siemens, and Nexuni. NVIDIA describes applications spanning server manufacturing, industrial workflows, and laundry handling. These are early testing signals, with deployment reliability still to establish. NVIDIA’s announcement details the tools and partners.
GR00T 2 gives robotics a concrete proposition to investigate: human video might supply transferable experience, while a small amount of robot-specific teaching closes the execution gap.
The next convincing result will come when another team can access the model, understand its terms, reproduce the evaluation, and show how often a person still has to step in.
Preferably while the camera keeps rolling.
Related reading
- NVIDIA and Hugging Face Are Building the GitHub Moment for Robots — Why shared models, datasets, and tools matter to robotics’ ability to scale.
- This Home Robot Learned to Do Your Dishes from 10 Million Real Family Routines — Another approach to collecting human experience without relying entirely on robot teleoperation.
