NVIDIA GR00T 2: Robots Learn Dexterous Skills From Human Video

NVIDIA unveiled GR00T 2 on November 9, 2026, at the Conference on Robot Learning (CoRL) in Austin, Texas, and the headline number is hard to ignore: a robot that had never practiced a task completed 40 of 50 unseen jobs at least once, with no demonstrations on that robot at all. GR00T 2 is a world action model, built on NVIDIA Cosmos, that learns dexterous manipulation by watching people work — and it arrives with a rebuilt training framework, Isaac Lab 3.0, and two new industrial benchmarks. The catch, and it matters, is that you cannot download the model yet.

That gap between announcement and availability is the story of this launch. The dataset behind it is already open, the training framework is already free, and the numbers NVIDIA published are unusually specific. What follows is what GR00T 2 does, what it was trained on, what it costs to try, and what NVIDIA has not shown yet.

What GR00T 2 actually is

Most robot models map what a camera sees straight to motor commands. GR00T 2 does something different: it predicts the future and the action together. One diffusion transformer reads recent video, the instruction, and the robot's current state, then denoises a short chunk of future video and the motor commands that would produce it, in a single pass. Small per-robot output heads turn that plan into commands for each body: parallel grippers, five-fingered dexterous hands, or a full humanoid.

The idea is that predicting how a scene should change teaches the model physics that a pure action predictor never sees. NVIDIA calls this a world action model, and it is the direct successor to the GR00T N1 series of robot foundation models. The practical payoff is skill transfer: a task demonstrated only by a human, or only with a handheld gripper, can carry over to a robot that never practiced it. In NVIDIA's tests, a two-handed robot lined a lab funnel with filter paper in 65% of trials after seeing only human demonstrations, and a YAM arm assembled a toy dump truck in 60% of trials after seeing the task only through a handheld UMI gripper.

The numbers that matter

NVIDIA evaluated GR00T 2 on 50 tasks that appear in human video or on other robots, but never in the target robot's own data. The target is a bimanual robot with two Sharpa hands. The results, all company-run, are the core of the announcement:

Two details deserve attention. First, the 95% figure comes with five demonstrations and a short fine-tune, not from the pretrained model alone (zero-shot is 40%). Second, the fine-tuning cost is a single consumer GPU for half an hour, which makes the few-shot claim usable for a small lab. The benchmark standing is NVIDIA's own claim, and no independent evaluation exists yet.

What it was trained on: 402,730 hours of people working

Human video is by far the largest part of GR00T 2's training mixture. The full mix, published on NVIDIA's research page, is 402,730 hours of egocentric video of people at work, 10,184 hours of demonstrations with a handheld UMI gripper across 2,897 tasks, and 23,500 hours of robot teleoperation spanning YAM and Franka arms, Sharpa and Wuji dexterous hands, and the Unitree G1 humanoid.

The human video comes from NVIDIA's Human-1M dataset, a collection of one million hours of egocentric footage. A 100,000-hour slice is openly available on Hugging Face as PhysicalAI-Human-1M-100K, covering 3,405,573 episodes of economically useful work, from sewing and brake repair to dumpling-making, dental work and tomato harvesting, and the dataset card states it is ready for commercial and non-commercial use. That open slice is the one part of the GR00T 2 stack you can actually download today.

NVIDIA also measured how much the human video helps. Pretraining four models on none, 1%, 10% and 100% of it, then testing on 72 downstream Dexter tasks, task progress rose at every step: from 64% to 77% in video rollouts and from 50% to 70% on the real robot. The scaling curve is the argument for the whole approach: more human video, better robot. It is also why NVIDIA open-sourced 100,000 hours of it.

GR00T 2 Turbo: the version that runs on the robot

The full model is too heavy to run onboard a humanoid, so NVIDIA built a smaller variant. GR00T 2 Turbo uses a 2-billion-parameter transformer and a higher-compression video encoder, and it steers a Unitree G1 humanoid through SONIC, a whole-body controller, with each prediction taking 124 milliseconds on the robot's own Jetson Thor computer (38 milliseconds on an RTX 5090). It plans 1.6-second chunks of whole-body motion about once per second, and SONIC executes them at 50 Hz.

On the G1, Turbo scooped and served popcorn in all 48 trials and cleared a table of diverse objects in 93% of trials. It also beats the full model on two YAM tasks (90% on fixed-pipe fittings and 100% on inserting a wire into a breaker) while trailing it on the other four YAM tasks and on five of six Dexter tasks. Adding simulated data lifts Turbo's out-of-distribution success from 6.7% to 63.3%, against 40.0% for GR00T 1.7 with the same treatment. The message is that the small model is not a cut-down demo: for onboard work, it is the model that matters.

Isaac Lab 3.0: the training ground, rebuilt

Alongside the model, NVIDIA announced Isaac Lab 3.0, the latest version of its open-source robot-learning framework. One timing detail is worth stating plainly: version 3.0.0 was published on GitHub on October 20, 2026, and is generally available. CoRL on November 9 was the showcase, not the release date.

The rebuild centers on Newton, a Warp-native physics engine that lets Isaac Lab model rigid bodies, deformable objects such as cables and cloth, and particles, plus complex articulated systems like multi-joint robotic hands. Researchers can switch between Newton and PhysX solvers, choose renderers for perception-based training, and use agents to set up training environments with simulation-ready assets. A new reinforcement-learning post-training library applies to foundation models like GR00T. In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton reached up to 2.5 times the training throughput of version 2.3 with PhysX, according to NVIDIA.

Version 3.0 also drops the requirement to run the full Isaac Sim stack: one task API works across physics, rendering and visualization backends, with kit-less execution through Newton and the OV libraries. Isaac Sim remains the complete path only when a workflow needs the full PhysX and RTX stack, ROS bridge, XR or livestreaming. The release targets Isaac Sim 6.2, Python 3.12, PyTorch 2.11 with CUDA 13.0, NVIDIA Warp 1.17, Newton 1.6.1 and RSL-RL 5.5.1. It is free on GitHub at isaac-sim/IsaacLab, and it is the piece of this launch any robotics team can start using immediately, unlike the model itself, which has no public weights yet.

Two new benchmarks for industrial robots

NVIDIA also released DexBench and VABAR, two open benchmarks designed to measure dexterity and adaptability on industrial tasks rather than tabletop demos. DexBench, developed by RLWRLD with NVIDIA, runs 15 tasks across four scenes (retail, logistics, manufacturing and hospitality), testing generalization across object, spatial and physics variations. Teams can run it on Manifold, Bifrost's evaluation platform, with no setup required and free credits for initial evaluations; Genesis AI and mimic robotics are among the early-access evaluators.

VABAR, the Variational Automation Benchmark for Agentic Robotics from UC Berkeley and NVIDIA, evaluates robots across six industrial tasks with standardized variations, difficulty tiers and success metrics. Both run on Isaac Lab-Arena, NVIDIA's open framework for scaled policy evaluation, whose developer preview of version 0.4 arrived at CoRL. The point of the pair is repeatability: policies run across thousands of variations in parallel, and the analysis shows which conditions cause failures, so the next training run knows what to fix.

Agentic robotics: agents that write the robot's program

The most forward-looking part of the announcement is not a model but a workflow. In agentic robotics, a frontier reasoning agent breaks a long task into steps, calls GR00T 2 as a tool for the contact-rich manipulation, watches for failures, and repairs the scene with code-as-policy skills when something drops or stalls. UC Berkeley's Graph-as-Policy framework connects individual skills into robot programs this way, testing them in simulation and refining them on failures — and Ambi Robotics, Bosch Research and Siemens are all exploring it for industrial automation.

The numbers NVIDIA published for this approach are striking. With an agent in the loop, GR00T 2 assembled a toy plane, a toy tractor and a Hot Wheels car in 98, 97 and 97 of 100 trials (292 of 300 in total), against 71 to 84 with the same policy and no agent, and 23 to 69 with supervised post-training alone. Inserting five Ethernet cables into a switch succeeded in 19 of 20 runs, and plugging five chargers into pointed-at outlets in 20 of 20, both well above the same tasks without reinforcement learning. Visual steering — a planner marking grasp and release points on the robot's camera view — raised short-horizon instruction following from 40% to 92% on YAM and from 30% to 70% on Dexter, and turned trash sorting on YAM from zero completed episodes to 17 of 20.

Two supporting techniques arrived with it. Success-Guided Sampling, from the University of Washington and NVIDIA, focuses practice on tasks a policy succeeds at about half the time, and is designed to scale training beyond a million parallel environments. AlphaNIST, built in Isaac Lab, trains one policy across all 20 tasks of the NIST Assembly Task Board and reaches over 98% training success against 81% for separate per-task policies, using roughly 39 times less GPU time.

Who is already using it

The partner list NVIDIA named is industrial, not academic. Foxconn is applying GR00T 2 to factory automation for server manufacturing, Siemens is evaluating it across industrial use cases, and Nexuni is testing it for robotic laundry handling. MicroAGI is using GR00T 2, post-trained through reinforcement learning, to swap Ethernet cables with a YAM robotic arm, automating the creation and calibration of its simulation environments with Isaac Lab 3.0 and Newton.

Others are building on the surrounding stack. Skild AI announced STEEL, a rigid-body physics solver built on Newton. Flexion showed Reflect, a robot-intelligence model trained with Isaac Lab and Newton. FieldAI is using the stack to train policies for unstructured environments, Dyna Robotics is training whole-body controllers for its semi-humanoid Taku robot, and the University of Texas at Austin is teaching robotic skills in simulation before students have hardware. NVIDIA also released a Video to Data workflow that turns human video into simulated robot demonstrations, and its EgoHand model, which reconstructs 3D hand motion from video, passed 40,000 Hugging Face downloads in the two weeks since launch.

What it costs, and what you can download today

There is no per-token price for GR00T 2, because there is no hosted API to buy. It is a model, not a service, and the cost of using it is hardware, time and, for now, patience.

The license question is the one to watch. GR00T 2's predecessor, GR00T N1.7, ships under the NVIDIA Open Model License with commercial use allowed, and its code is Apache 2.0. But NVIDIA has not stated any license for GR00T 2, and the weights are not out. Do not assume the new model carries the old model's terms until NVIDIA publishes them. The Jetson Thor price above comes from a November 9, 2026 price check of NVIDIA's Jetson FAQ, which lists the developer kit at $5,499 after the July 2026 increase from its $3,499 launch price; the pound and Canadian dollar figures are approximations at November 10, 2026 market rates (1 USD = 0.745 GBP, 1.41 CAD).

GR00T 2 vs GR00T N1.7

The comparison that matters most is against the other open robot stack. Hugging Face's LeRobot 0.7.0, released November 5, adds world action models and vision-language-action models of its own, with remote GPU inference and streaming training on nearly 300TB of open robot data. NVIDIA's bet is different: far more human video, a simulation stack built for industrial evaluation, and an agentic layer on top.

The fine print

Every benchmark number in this article is NVIDIA's own. There is no independent evaluation of GR00T 2 yet, and the 50-task evaluation is a company-chosen set on company hardware. The 95% figure needs five demonstrations and a fine-tune per task; the pretrained model alone sits at 40%. The "top three" benchmark claim covers four specific leaderboards NVIDIA named.

The research page is candid about current limits, and they are worth repeating. On Dexter, visual steering marks place more items in trash sorting but rarely complete the full task. Marks convey neither an object's final orientation nor the motion between grasp and release (pouring, wiping), which still comes from language. And the "open" label needs the qualifier above: the dataset and the training framework are open today, while the model weights and code are not. Isaac Lab 3.0 is available now; GR00T 2 is an announcement with a December code date and no weights date at all.

None of that makes the launch thin. A robot model that learns a new dexterous skill from five human demonstrations and a 30-minute fine-tune on one consumer GPU changes who can teach robots, and the industrial partner list (Foxconn, Siemens, MicroAGI) suggests the first users will be factories, not living rooms. But the honest version is two-part: a strong set of results, and a download page that is still mostly "coming soon."

Common mistakes if you try it

Do not wait for the weights to start. Isaac Lab 3.0 and the 100,000-hour Human-1M slice are available now, and both are the inputs a GR00T 2 workflow needs: the framework to train in, and the human video to learn from. Do not assume GR00T 2's license matches N1.7's; check the terms when NVIDIA publishes them, before building a product on it. Do not read the 95% success rate as production-ready: it is a 50-task lab evaluation, and long-horizon tasks with orientation changes still fail. If you install Isaac Lab 3.0, note the pinned stack (Python 3.12, PyTorch 2.11, CUDA 13.0), because the framework moved on from the 2.x requirements.

When can I download GR00T 2?

Not yet. As of November 10, 2026, no GR00T 2 weights exist on Hugging Face, and the research page lists code as "Available in December," according to MLQ's read of the page. The 100,000-hour Human-1M dataset and Isaac Lab 3.0 are available now.

Is GR00T 2 free?

There is no price, because there is no hosted service to buy and no public download yet. NVIDIA calls it an open frontier world action model, but no license has been stated, so commercial use cannot be assumed until NVIDIA publishes terms.

What hardware do I need to try GR00T 2?

For fine-tuning a new task, NVIDIA's figure is a single RTX 5090 for 30 minutes per task. For running the Turbo variant onboard a humanoid, the reference computer is Jetson Thor, whose developer kit lists at $5,499 (about £4,100 / CA$7,750 at November 10, 2026 rates).

How is GR00T 2 different from GR00T N1.7?

GR00T 2 is a world action model: it predicts future video and robot actions together in one diffusion transformer, which lets it learn skills from human video alone. N1.7 is a vision-language-action model, and it is the version available today, under the NVIDIA Open Model License with commercial use allowed.

What is Isaac Lab 3.0?

It is NVIDIA's open-source robot-learning framework, generally available since October 20, 2026. Version 3.0 adds the Newton physics engine for deformable objects, a single task API across backends, and up to 2.5 times the training throughput of version 2.3 in NVIDIA's dexterous manipulation benchmark.

Does GR00T 2 work on any robot?

Not out of the box. It uses small per-robot output heads, and NVIDIA demonstrated it on YAM and Franka arms, Sharpa dexterous hands and the Unitree G1 humanoid, with skills transferring between bodies. A new robot still needs adaptation data: five demonstrations per task in NVIDIA's tests.

GR00T 2 is a clear sign that robot learning is shifting from teleoperation to human video: watch people work, and the robot learns the job. The results are real, the dataset is open, and the training framework is free today. The model itself is the missing piece, and December is when we find out whether NVIDIA ships it the way it shipped the data. Until then, Isaac Lab 3.0 and 100,000 hours of people doing useful work are the parts worth your time. Try the framework, browse the dataset, and check back when the weights land.

Sources

• NVIDIA Blog: "Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation" (November 9, 2026)
• NVIDIA Research: "GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation"
• GitHub: isaac-sim/IsaacLab (v3.0.0, published October 20, 2026)
• Hugging Face: nvidia/PhysicalAI-Human-1M-100K dataset
• The Robot Report: "NVIDIA launches GR00T 2 and Isaac Lab 3.0 at CoRL" (November 9, 2026)
• MLQ: "Nvidia launches GR00T 2, a robot model that learns new skills from human video, and Isaac Lab 3 at CoRL" (November 9, 2026)
• AI Weekly: "Nvidia debuts GR00T 2 robotics model, 71% on 50 tasks" (November 10, 2026)
• DexBench: dexbench.ai
• Physical AI Lab: "NVIDIA Jetson Thor Price: $5,499 Kit, T5000 $4,999, T4000 $2,999" (prices re-checked November 9, 2026)