By RWT Robotics Editorial Team Published
CONFIRMED
NVIDIA announced GR00T 2 at the Conference on Robot Learning (CoRL) in Austin, Texas, The Robot Report reports. The headline numbers are impressive, but they come from NVIDIA alone, and coverage notes the company has not said when developers can get the model.
What we know
NVIDIA’s blog, as reported by The Robot Report, describes GR00T 2 as an open world action model built on NVIDIA Cosmos. It is trained on both robot data and human data. The human portion includes about 100,000 hours of egocentric human video, meaning first-person footage of people doing things with their hands.
NVIDIA says the model can learn skills from human demonstrations and transfer those skills across different robot bodies. That second part matters most. Cross-embodiment transfer is one of the hard, unsolved problems in robot learning. A human hand, a parallel gripper and a five-fingered humanoid hand do not share kinematics, and mapping one to another is where many promising approaches have stalled.
NVIDIA also gave numbers. The company says GR00T 2 completed 40 of 50 tasks at least once without any task-specific demonstrations. It says the model averaged 71% success with one demonstration per task and 95% with five.
These are NVIDIA’s claims, graded as company statements. We found no independent verification. We also have not established whether the demonstrations behind these results were autonomous, scripted, sped up or teleoperated. Those details decide how much the numbers mean.
The “at least once” phrasing in the zero-demonstration figure also needs care. Completing a task once across some number of attempts is a much lower bar than completing it reliably. The 71% and 95% figures are averages, and in the material we reviewed they have no stated trial counts, task difficulty breakdowns or failure modes.
Isaac Lab 3.0 is better defined. NVIDIA’s blog and The Robot Report describe it as NVIDIA’s open-source robot learning framework. It lets agents set up training environments using simulation-ready assets. It can switch between the PhysX and Newton physics solvers, and it adds a reinforcement post-training library.
According to the NVIDIA Blog, early users testing GR00T 2 are Nexuni, Foxconn, Siemens and MicroAGI. Their use cases are laundry handling, server manufacturing, industrial applications and cable swapping. Laundry and cable work are good stress tests because both involve deformable objects, which remain hard for robot manipulation.
This announcement fits a pattern. At CoRL 2024 in Munich, NVIDIA announced Isaac Lab 1.2 and new Project GR00T humanoid learning workflows. At CoRL/Humanoids 2025 in Seoul, it announced the Newton physics engine beta, said GR00T N1.6 was coming soon and previewed Isaac Lab 2.3. Newton, an open-source physics engine, now appears as a solver option in Isaac Lab 3.0.
What we don’t know yet
The central question for this piece has no answer yet: how much of GR00T 2’s learning from human video shows up on real hardware, and how much in simulation? None of the material we reviewed says whether the 40-of-50 result or the 71% and 95% averages came from physical robots, simulated robots or a mix. NVIDIA sells both the simulator and the model, so the distinction is not academic. Sim-to-real transfer is its own source of failure, separate from human-to-robot transfer.
We also do not know which robot bodies were used in the evaluations. A claim of cross-embodiment transfer is only as strong as the variety of embodiments tested, and we have no list.
We do not know what the early users have actually achieved. NVIDIA names the partners and their target tasks. It does not give results from those deployments, and “testing” can mean anything from a lab pilot to a production line.
Most practically, we do not know when or how developers can use GR00T 2. Coverage notes that NVIDIA gave neither a release date nor an access method. The model is called “open,” but the terms of that openness have not been set out.
Isaac Lab 3.0’s timing is also unclear. One secondary report says it became generally available in late October 2026, before the CoRL announcement. We could not confirm that from an official source, so treat it as unverified.
What it means
The idea behind GR00T 2 is the right one. Robot data is expensive and slow to collect. Human video is plentiful. A model that turns 100,000 hours of people folding, plugging and assembling into transferable robot skills would change the economics of robot learning. That is why this direction gets so much attention across the field.
But a direction is not a result. The few-shot numbers, 71% with one demonstration and 95% with five, would be strong if they hold on physical robots across varied hardware. We cannot yet tell whether they do.
Isaac Lab 3.0 is the more concrete deliverable. It is open source, its features are specific, and swapping physics solvers is useful for anyone checking whether a policy depends on one simulator’s quirks.
For developers, the honest summary is short. The framework is described in detail. The model is announced but not dated. Until NVIDIA publishes access terms and evaluation details, or until partners like Foxconn and Siemens report results, GR00T 2 is a promising claim, not a tool you can plan around.
Keep reading
Robotics NewsBoston Dynamics puts wheels on Atlas to test whole-body reinforcement learningBoston Dynamics' 'Atlas On a Roll' post shows wheeled Atlas as a whole-body RL experiment. It is research,…Nov 9, 2026 · 4 min read
Robotics NewsNvidia expands Korea physical AI work with Samsung, Hyundai, Doosan and SK hynixNvidia named Samsung, Hyundai, Doosan and SK hynix in its Korea physical AI update. Stakes in four startups,…Nov 9, 2026 · 5 min read
Robotics NewsMachenta’s $40M seed is confirmed, but one-demo robot learning is unproven outside the labImperial says Machenta raised a $40M seed led by Creandum and USV. One-demo learning from simulation is the…Oct 27, 2026 · 4 min readCredits and sources
By RWT Robotics Editorial Team
- Published
