What Is NVIDIA GR00T 2? The Robot Model That Learns Dexterous Skills From Human Video

GR00T 2 is NVIDIA’s new foundation model for robots — a system that learns to handle objects by watching videos of people doing the same things. Introduced on Monday 9 November 2026 at the Conference on Robot Learning (CoRL) in Austin, Texas, it tackles one of the oldest problems in robotics: a robot that walks into a new situation and figures out, with little or no extra training, how to use its hands.

In short: instead of programming a robot’s movements, or steering it through a task hundreds of times, GR00T 2 studies first-person video of humans — 100,000 hours of it, released openly — and learns the general skill behind the movement. One example of a new task brings about 70% success; five bring about 95%, according to NVIDIA’s own tests. This guide explains what the model is, how the human-video trick works, what those numbers do and do not prove, and who should care.

What is NVIDIA GR00T 2?

GR00T 2 is what NVIDIA calls a “world action model”. Most robot brains are vision-language-action models: they look at a scene, read an instruction in plain language, and output motor commands. A world action model does one thing more — it predicts what the scene will look like after it acts. Before moving, it imagines the next few moments of video and the actions that produce them, together, in a single diffusion transformer. The prediction is the plan.

The model is built on NVIDIA Cosmos, specifically on Cosmos-Predict2.5-2B. It takes recent camera views, the robot’s state and a written instruction, and generates the expected future views plus the next actions. Small per-robot output heads adapt the same core model to different bodies, so one pretrained checkpoint covers parallel-gripper arms, dexterous hands and a humanoid. A faster Turbo variant uses a 2-billion-parameter transformer and predicts in 124 milliseconds on NVIDIA’s Jetson Thor robot computer — fast enough, in principle, for real-time control.

GR00T is pronounced “groot”, and the line descends from NVIDIA’s Project GR00T, announced in 2024 as a foundation model for humanoid robots. GR00T 2 succeeds the GR00T N1 series: the earlier N1.6 and N1.7 models were 3-billion-parameter vision-language-action systems on a Cosmos backend. The name change — from vision-language-action to world action model — signals the new approach: predicting the future, not just reacting to the present.

How learning from human video works

The core problem GR00T 2 attacks is data scarcity. Robot data is expensive: every hour of demonstration needs a physical robot, a human operator and a teleoperation rig, and only covers that one robot body doing that one task. Human video is the opposite — cheap, abundant, and full of exactly the skills robots lack: picking things up, folding, pouring, assembling, using tools. The catch is that a video of a person is not a video of a robot: hands are not grippers, and a robot cannot copy a movement it cannot perform.

NVIDIA’s answer is to train on both, at scale, and let the model learn the mapping. The pretraining mix combines 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation. The handheld gripper is the bridge: a person holds a gripper tool and performs tasks the way a robot would. NVIDIA reports that adding the human video raised real-robot task progress from 50% to 70% across 72 tasks — the clearest evidence the video is doing real work.

Of the human video, 100,000 hours are being released openly as the Human-1M-100K dataset: video, language annotations and head-and-hand motion data, under a CC BY 4.0 licence that permits commercial use. The dataset is useful even to people who never touch GR00T 2 — any lab building robot learning systems can train on it. Two companion tools are open as well: the Video to Data workflow, which converts human video into simulated robot demonstrations, and EgoHand, an open model that reconstructs 3D hand motion from video, with over 40,000 Hugging Face downloads in two weeks.

Once pretrained, the model adapts to a new task with a handful of demonstrations: five of them — under three minutes of robot data per task — followed by about 30 minutes of fine-tuning a small adapter on a single RTX 5090 graphics card. The model arrives already knowing roughly how hands work; the demonstrations only teach it the specifics of this task, on this robot.

What the 50-task results do and do not prove

The headline numbers come from 50 tasks the model had not been trained for. NVIDIA reports GR00T 2 completed 40 of them at least once with no task-specific demonstrations, averaging 40% success. With one demonstration per task, average success rose to 71%; with five, it reached 95%. Each number measures something different.

The 40% zero-shot figure is the most honest measure of generalisation: put this model in front of a task it has never seen, with no example at all, and it succeeds two times in five — far from reliable, and far ahead of what robot models could do a few years ago. The 71% one-shot figure is the practical one: a single demonstration is cheap, and it more than doubles the chance of success. The 95% five-shot figure is the least informative. NVIDIA’s whitepaper notes the five-demonstration tasks were selected from among those the model could already learn with five examples at 80% success or better. The average is high partly by construction — it is not “GR00T 2 succeeds 95% of the time on new tasks”.

The cross-body results are more revealing than any average. NVIDIA reports 65% success on a filter-paper setup learned purely from human demonstrations, with no task-specific teleoperation, and 60% on a toy dump-truck assembly transferred from a handheld gripper to a YAM robotic arm. Those tests measure what matters most for real deployment: whether a skill learned from one kind of body works on another. A model that only works on the robot it trained on is a demo; one that transfers across bodies is a tool.

Beyond the 50 tasks, GR00T 2 ranks among the top three models across the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins, according to NVIDIA. With corrective data and a reasoning agent running skills in sequence, the robot completed 98 toy-plane, 97 toy-tractor and 97 Hot Wheels assemblies in 100 trials each — against 69, 51 and 23 with supervised fine-tuning alone. NVIDIA’s whitepaper also reports roughly 70% average success on long-horizon post-training benchmarks. All of these figures are vendor-reported from controlled tests, and none has been independently reproduced — the model is not downloadable until December, so outsiders have had no chance to check them.

Isaac Lab 3.0 and the training pipeline

A pretrained model is only half the story. The other half is simulation: where robots practise, fail and improve without breaking anything. NVIDIA’s simulation framework is Isaac Lab, and its latest release, Isaac Lab 3.0, was the other half of the CoRL announcement. It lets researchers use AI agents to set up training environments quickly, switch between the PhysX and Newton physics solvers, choose renderers for perception-based training, and apply a new reinforcement-learning post-training library to foundation models such as GR00T.

The important addition is Newton, an open-source physics engine. Accurate physics is what lets a simulated skill survive contact with the real world: Newton models deformable objects — cables, cloth — and complex articulated systems such as multi-jointed robotic hands, where older simulators diverge. In a vision-based dexterous manipulation benchmark, NVIDIA reports Isaac Lab 3.0 with Newton achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX. One clarification: Isaac Lab 3.0 itself has been generally available since 20 October 2026 — the CoRL announcement pairs it with GR00T 2; it does not launch it.

The pipeline has three layers. Pretraining gives the model its broad skills. Reinforcement-learning post-training sharpens them: MicroAGI is using GR00T 2 post-trained this way to swap Ethernet cables with a YAM robotic arm. An agentic layer sits on top, with a reasoning agent sequencing skills, monitoring progress and invoking recovery routines when parts drop or work stalls. The agent layer works, but it costs time: in visually steered trash sorting, the robot waited for the agent during a median 38% of an episode. Cutting that waiting while keeping the reliability is an open problem.

Evaluation got its own upgrades. Isaac Lab-Arena 0.4, in developer preview, runs policy evaluations across thousands of variations in parallel, with failure analysis showing which conditions cause failures. Two new open benchmarks, DexBench and VABAR, are among the first designed to measure robot dexterity and adaptability across industrial tasks. DexBench, from RLWRLD with NVIDIA, has 15 tasks across retail, logistics, manufacturing and hospitality scenes; VABAR, from UC Berkeley and NVIDIA, covers six industrial tasks with standardised variations and difficulty levels. In a simulation demo, AlphaNIST trained a single policy on all 20 NIST Assembly Task Board tasks — parts as small as 4 millimetres — reaching over 98% training success with roughly 39 times less GPU time than separate per-task policies.

Where GR00T 2 is being trialled

The model is already in the hands of industrial users — but that means testing and evaluation, not production. Nexuni is testing GR00T 2 for robotic laundry handling. Foxconn is applying it to factory automation for server manufacturing. Siemens is evaluating it across industrial use cases. MicroAGI’s Ethernet cable swapping, described above, is the most concrete public trial. No production rollout has been announced by any of them.

The wider ecosystem is moving with it. Skild AI announced STEEL, a rigid-body physics solver built on Newton to accelerate robot training with Isaac Lab. Flexion showed how Isaac Lab and Newton help train Reflect, its robot-intelligence model combining reasoning and whole-body manipulation. FieldAI is using Newton, Isaac Lab and Isaac Sim to train and validate robot policies for unstructured real-world environments. Dyna Robotics will demonstrate parallel simulation in Isaac Lab training whole-body controllers for its semi-humanoid Taku robot. UC Berkeley’s Graph-as-Policy framework uses agents to connect individual skills into interpretable robot programs, rehearsing them in simulation and revising after failures — ten rehearsal iterations lifted its Make Popcorn result from 33% to 94% in simulation and 90% on a real robot. For another agent-style launch from the same period, see our guide to Tencent’s Hy4 computer-use model.

Open-access status and the December code release

GR00T 2 is introduced as an open model, but “open” needs careful reading. Genuinely open today: the 100,000-hour Human-1M-100K dataset under CC BY 4.0, the Video to Data workflow, the EgoHand model, Isaac Lab 3.0, the Newton physics engine, and the DexBench and VABAR benchmarks. Not open yet: the model itself. NVIDIA’s research page lists code as “Available in December” and states no licence or release date for the model weights. Until December, researchers can read the paper and use the dataset, but they cannot run GR00T 2 themselves.

That gap matters for two reasons. First, the benchmark numbers cannot be independently checked until the code and weights are available — every figure in this guide is NVIDIA’s own, from controlled tests. Second, the licence for the weights is still unstated. The dataset’s CC BY 4.0 licence permits commercial use, but whether the model weights will carry the same permissive terms or a research-only licence is not yet known. The December release is the event to watch: its timing, its licence, and whether outside labs can reproduce the 50-task numbers.

How GR00T 2 compares

Against its predecessor, the change is architectural. The GR00T N1 series were vision-language-action models: they saw, they read an instruction, they acted. GR00T 2 adds the world-model step — predicting the future it is about to create — and trains on a far larger pool of human video. The N1 models were 3-billion-parameter systems; GR00T 2’s Turbo variant uses a 2-billion-parameter transformer, smaller but, on NVIDIA’s numbers, more capable.

Against other robot foundation models, the comparison is harder to draw cleanly, because the field is young and every lab benchmarks differently. The broad pattern: most robot foundation models are trained mainly on teleoperation data — a human steering a robot — which is expensive and ties the model to one body. GR00T 2’s bet is that human video, at 400,000-plus hours, is a better foundation, and that predicting future video as well as actions makes the model more robust. The early evidence supports the bet: the 50% to 70% jump in real-robot task progress when human video was added, and the cross-body transfer results, suggest a real change in approach rather than a bigger version of the old one. The counter-evidence is that no independent lab has yet run the model, and the strongest numbers come from tasks the model was already good at. For another model from the same week that takes a different approach — teaching software agents to use computers by sight — see our guide to Gemini Omni 1.2 Flash, Google’s 15-second video model.

Who should use it — and who should wait

GR00T 2 is most relevant right now to robotics researchers and teams building manipulation skills: the dataset, the simulation tools and the benchmarks are available today, and they are useful with or without the model. Factory-automation teams evaluating dexterous work — cable handling, assembly, laundry, sorting — should follow the industrial trials, the first real-world tests of whether the approach survives contact with production. Students of robotics should read the paper: the world action model idea and the human-video training recipe are likely to shape the next few years of the field.

Others should wait. If you need a model to download today, GR00T 2 does not exist yet — December is the earliest date, and the licence is unstated. If you need proven production reliability, the trials are evaluations, not deployments, and the 40% zero-shot average is a reminder that the model still fails often without examples. And if you need unsupervised operation, the agent layer’s latency — the robot waiting for its overseer during a median 38% of an episode — shows the autonomy is not yet seamless.

What to watch next

First, the December code release: whether it arrives, what licence the weights carry, and whether outsiders can reproduce the 50-task results once they can run it. Second, the industrial trials — whether Nexuni, Foxconn and Siemens move from evaluation to deployment, and what failure modes they report. Third, the benchmarks: DexBench and VABAR are designed to measure exactly the industrial dexterity GR00T 2 claims, and their first independent results will be a better test than any vendor number. Fourth, the wider field: if human video becomes the standard training recipe, expect rival labs to release their own video-trained robot models within months.

For now, GR00T 2 is best understood as a bet with strong early evidence: that the fastest way to teach a robot to use its hands is to let it watch people use theirs. The dataset is open, the approach is backed by a large measured gain, and the industrial trials are real. Whether the bet pays off will be decided in December, when the model itself becomes available to check.

Related news coverage: our news report on the NVIDIA GR00T 2 launch