NVIDIA GR00T 2 Robot Explained: Learns Dexterous Skills From Human Video, 95% Success and December Open Release
Illustrative image: Unitree's G1 humanoid robot. GR00T 2 is a software model, not a robot — the G1 is one of the robot bodies whose teleoperation data trained it. Image: Sayanesy, Wikimedia Commons, CC0.
📌 Key Highlights
- NVIDIA unveiled GR00T 2 on 9 November 2026 at the Conference on Robot Learning (CoRL) in Austin, Texas — an open frontier world action model built on NVIDIA Cosmos.
- It learns dexterous manipulation from human video: its pretraining mix includes 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations and 23,500 hours of robot teleoperation.
- In NVIDIA's own tests, GR00T 2 completed 40 of 50 unseen tasks at least once with zero demonstrations, averaged 71% success with one demonstration per task, and 95% with five.
- The 100,000-hour Human-1M-100K dataset is openly available now under a CC BY 4.0 licence; the model weights and a lighter Turbo variant are slated for an open release in December.
- Isaac Lab 3.0, the open-source simulation framework behind the training, is generally available — with up to 2.5x faster training than its predecessor.
What happened?
At this week's Conference on Robot Learning in Austin, Texas, NVIDIA introduced a package of open robotics technologies — and the headline item is GR00T 2, the next generation of its robot foundation model. The conference runs 9–12 November, and NVIDIA's robotics chief Jim Fan presented the model on the opening day. Two NVIDIA papers, GRAIL and RoboTTT, are also among the six finalists for the conference's Best Paper Awards.
The announcement marks a shift in what robot learning is trying to solve. Simulation has long been the proving ground for robots learning to walk, roll and hop. GR00T 2 targets the harder problem: fine manipulation — aligning gears, connecting cables, assembling small components — the fiddly work that has kept robots out of most factory and household jobs. NVIDIA says the model is already being tested by partners: Nexuni for robotic laundry handling, Foxconn for server-manufacturing automation, Siemens across industrial use cases, and MicroAGI for Ethernet cable swapping with a YAM robotic arm.
NVIDIA GR00T 2 robot explained: how a world action model learns
A world action model predicts what happens next in the physical world and what actions a robot should take, together, in one model. GR00T 2 is a 2-billion-parameter transformer built on Cosmos-Predict2.5-2B: it takes in recent camera views, a written instruction ("pick up the cable and plug it in") and the robot's current state, then generates a short chunk of predicted future video plus the motor commands to make it happen — imagining the scene changing as it acts, then executing the chunk and repeating. Small per-robot output heads turn that plan into motor commands for each body, so one checkpoint covers gripper arms, dexterous hands and a humanoid robot.
| GR00T 2 at a glance | Detail |
|---|---|
| What it is | Open frontier world action model for robots |
| Built on | NVIDIA Cosmos (Cosmos-Predict2.5-2B), 2-billion-parameter transformer |
| Training data | 402,730 hours of first-person human video; 10,184 hours of handheld-gripper demos across 2,897 tasks; 23,500 hours of robot teleoperation |
| Zero demonstrations | 40 of 50 unseen tasks completed at least once |
| One demonstration per task | 71% average success |
| Five demonstrations per task | 95% average success |
| Open dataset | Human-1M-100K — 100,000 hours of human video, CC BY 4.0, available now |
| Model weights | Slated for open release in December 2026, with a lighter GR00T 2 Turbo variant |
The big idea is the training data. Teaching a robot the traditional way — teleoperation, where a human drives the robot remotely while it records every movement — is slow and expensive, because every new skill needs a robot, an operator and hours of demonstrations. GR00T 2 flips the balance: there are billions of hours of people doing things, and almost none of it needs a robot to record. The model learns the shape of a skill from watching us — how a hand approaches a handle, how a cable bends when pulled — then transfers that understanding to a robot body that has never practised the task. NVIDIA calls this cross-embodiment transfer: in its tests, a filter-paper task learned from human demonstrations reached 65% success, and a toy dump-truck assembly transferred from a handheld gripper to a YAM arm reached 60%.
Illustrative image: DLR's SpaceJustin, a humanoid torso built for dexterous manipulation research — the kind of fine hand work GR00T 2 is designed to learn. Image: DLR (Deutsches Zentrum für Luft- und Raumfahrt), Wikimedia Commons, CC BY 3.0 de.
Two honest caveats belong with the headline numbers. First, the 71% and 95% figures are NVIDIA's own test results: the 50 tasks are unnamed, no independent replication has been published yet, and the technical whitepaper selected tasks the model could learn with five episodes at 80% success or better — so the averages are high partly by construction. Second, NVIDIA calls GR00T 2 "open", but the weights, code and licence are not published yet. What is genuinely open today is the 100,000-hour dataset and the Isaac Lab simulation framework; the model itself arrives in December, when NVIDIA says the weights for GR00T 2 and the lighter Turbo variant will be released as open source, along with the training recipe. Turbo keeps the same 2 billion parameters but uses four times fewer spatial video tokens, so it can run on the robot — one prediction takes 124 milliseconds on NVIDIA's Jetson Thor computer.
What Isaac Lab 3.0 adds: the practice ground
A pretrained model gives a robot a core set of skills; simulation is where those skills are refined at scale before touching a physical robot. Isaac Lab 3.0, NVIDIA's open-source robot-learning framework, became generally available on 20 October 2026. Researchers can now use AI agents to set up training environments quickly, switch between the PhysX and Newton physics solvers, and apply a new reinforcement-learning post-training library to foundation models like GR00T. Tighter Newton integration lets the simulator model deformable objects such as cables and cloth, plus robot hands with many interacting joints — the exact things fine manipulation needs. In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton reached up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX.
The simulation-first approach is already showing results. AlphaNIST, built in Isaac Lab, trains a single policy across all 20 tasks of the NIST Assembly Task Board — manipulating millimetre-scale parts on a yes-or-no success signal — at more than 98% training success, against 81% for separate per-task policies, using roughly 39x less GPU time. Two new open benchmarks measure the same progress: DexBench (RLWRLD with NVIDIA) tests 15 tasks across retail, logistics, manufacturing and hospitality scenes, and VABAR (UC Berkeley and NVIDIA) evaluates robots across six industrial tasks. NVIDIA also released an open Video to Data workflow that turns human video into simulated robot demonstrations, and its EgoHand model — which reconstructs 3D hand motion from video — has passed 40,000 Hugging Face downloads in the two weeks since launch.
Why it matters for India readers
India has no shortage of robotics ambition — from factory automation to humanoid startups like the Embodied AI AIR humanoid, which learns factory jobs in a month — but robot training data and hardware have been the bottleneck. GR00T 2 attacks both sides of that bottleneck in ways Indian teams can actually use. The 100,000-hour open dataset is available now, free, under a licence that permits commercial use — a research lab or startup in India can start training on it today without owning a single robot. And Isaac Lab 3.0's simulation-first approach means teams can develop and test skills before physical hardware exists at all; NVIDIA notes the University of Texas at Austin already uses it to let students start developing before robots arrive, and the same logic applies to Indian institutes and startups. When the weights land in December, the whole stack — data, simulator, model — becomes something a small Indian team can build on rather than rent from a big lab.
That is also the right scale for expectations. No India-specific deployment of GR00T 2 has been announced, the partner trials at Nexuni, Foxconn and Siemens are tests and evaluations rather than production rollouts, and the model itself cannot be downloaded yet. The near-term value for Indian readers is the open dataset and the simulation tools — and the longer-term signal that robot learning is moving from "teach each robot each task" to "show the model a video and let it generalise", which is the shift that makes robotics affordable enough to build in India. The same agentic pattern is spreading elsewhere in AI, as our explainer on Google DeepMind's AI agents designing quantum programs shows.
Should you care?
FAQs
What is NVIDIA GR00T 2?
GR00T 2 is NVIDIA's open frontier world action model for robots, unveiled on 9 November 2026 at the Conference on Robot Learning. Built on NVIDIA Cosmos, it predicts what a robot's camera will see next and what actions to take, together, so robots can learn dexterous manipulation skills — the successor to the GR00T N1 series of robot foundation models.
How does GR00T 2 learn from human video?
Instead of relying only on robot demonstrations, GR00T 2 is pretrained largely on first-person video of people doing everyday tasks — 402,730 hours of it — plus handheld-gripper demonstrations and robot teleoperation. It learns the shape of a skill from watching humans, then transfers it to a robot body that has never practised the task, a capability NVIDIA calls cross-embodiment transfer.
Is GR00T 2 open source, and when can I download it?
Not yet. NVIDIA calls GR00T 2 open, but the model weights, code and licence terms have not been published. NVIDIA says the weights for GR00T 2 and a lighter Turbo variant will be released as open source in December 2026, along with the cross-embodiment training recipe. What is open today is the 100,000-hour Human-1M-100K dataset (CC BY 4.0) and the Isaac Lab 3.0 simulation framework.
What is Isaac Lab 3.0 and what does it add?
Isaac Lab 3.0 is NVIDIA's open-source robot-learning framework, generally available since 20 October 2026. It lets researchers train robots in simulation at scale — with AI agents setting up training scenes, the Newton physics engine modelling cables, cloth and multi-jointed hands, and up to 2.5x the training throughput of the previous version.
Takeaway
GR00T 2 is NVIDIA's bet that robots will learn the way chatbots did — from a vast pile of human data, not from one demonstration at a time. The December open release is the date that decides whether that bet pays off for everyone else.
If you build robots, the part you can use today is the data: Human-1M-100K is on Hugging Face now, and Isaac Lab 3.0 is open source — the model weights follow in December.
Sources
- NVIDIA Blog — Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation (9 November 2026)
- The Robot Report — NVIDIA launches GR00T 2 and Isaac Lab 3.0 at CoRL (9 November 2026)
- Unite.AI — NVIDIA Introduces GR00T 2 for Robots Learning From Human Video (9 November 2026)
- Humanoids Daily — "Robotics Is No Longer an Elitist Sport": NVIDIA Unveils GR00T 2 and Open-Sources the Stack Behind It (9 November 2026)
- TechToday — NVIDIA's GR00T 2 Explained: The Robot Brain That Learns From Watching Humans (10 November 2026)
- Derrington Collaborative AI — NVIDIA GR00T 2: robots that learn by watching people (10 November 2026)
- Artiverse — NVIDIA's Robot Strategy Moves Beyond Imitation Learning (10 November 2026)
Comments
Post a Comment