At the Conference on Robot Learning (CoRL) in Austin on November 9, 2026, NVIDIA quietly dropped one of the most significant robotics announcements of the year: GR00T 2, a world action model that learns physical tasks by watching people do them.
The idea sounds deceptively simple. Feed a robot enough video of humans performing a task — say, folding laundry or assembling a component on a factory floor — and it generalizes well enough to replicate those movements in new environments. The difference with GR00T 2 is the scale and the architecture.
What GR00T 2 Actually Is
GR00T 2 is a 2-billion-parameter world action model built on NVIDIA's Cosmos-Predict2.5-2B video model. Unlike traditional robot learning systems that separate perception from action, GR00T 2 generates both a short chunk of predicted future video and the corresponding motor commands simultaneously — in a single forward pass. The robot isn't just planning what to do; it's simulating what the world will look like as it does it.
Training used 100,000 hours of NVIDIA open-source human egocentric video data — footage shot from a first-person perspective, capturing how people naturally manipulate objects. That dataset is what makes few-shot transfer possible: the model has seen enough human motion to infer robot motion from just a handful of demonstrations.
The Numbers
NVIDIA's internal tests showed a 71% average success rate across 50 tasks when given a single demonstration. Jump to five demonstrations and that number rises to 95%. There is a faster variant — GR00T 2 Turbo — which runs inference in 124 milliseconds on the robot's own onboard hardware, without a cloud connection.
These are NVIDIA's own figures from their own test suite, which matters. Independent replications on external benchmarks haven't published yet, so the headline numbers should be read as promising internal results rather than settled benchmarks.
Who's Already Using It
Three industrial partners were announced alongside the model. Foxconn is testing GR00T 2 on server manufacturing lines, Nexuni is evaluating it for robotic laundry handling, and Siemens is exploring industrial applications. These are real production environments, not lab demos — which makes the deployment timeline more credible than most robotics press releases.
NVIDIA also released Isaac Lab 3.0 at CoRL, its GPU-accelerated simulation environment for training robots at scale. The idea is a closed loop: simulate in Isaac Lab, deploy on GR00T 2, measure on the new open benchmarks NVIDIA published alongside both tools.
The Openness Question
NVIDIA calls GR00T 2 "open," but the weights weren't available at launch. The company has said code will ship in December 2026, with no announced license for the model weights themselves. That's a meaningful asterisk — open-source robotics infrastructure tends to accelerate community research, and "open later" is a different value proposition than "open now." Whether the December release includes downloadable weights or just training tooling will determine how much of the ecosystem can actually build on it.
Why This Matters
The bottleneck in deploying robots into new tasks has always been data collection and re-training. If a model can generalize from five human demonstrations to a 95% success rate on a novel physical task, that changes the economics of robotics deployment dramatically. The gap between "needs a team of engineers to reprogram" and "show it five times and it works" is the gap between specialty automation and general-purpose robotics.
GR00T 2 isn't there yet across all tasks — 95% success with five demos under NVIDIA's conditions is a research milestone, not a production guarantee. But it's the kind of milestone that makes the next few years worth watching closely.