Nvidia introduced GR00T 2 at the Conference on Robot Learning (CoRL) on 9 November, calling it “an open frontier world action model built on NVIDIA Cosmos” in its CoRL blog post. Nvidia’s GEAR research page says the model is pretrained on 402,730 hours of egocentric video of people at work, 10,184 hours of handheld-gripper demonstrations and 23,500 hours of robot teleoperation.
By our arithmetic, human video is 92.3% of that 436,414-hour mix, and robot teleoperation is 5.4%. For every hour of robot data, the model saw about 17 hours of people doing the work.
Business Pill · Physical AI
A short explainer of physical AI, models that act through robot bodies. It teaches the general idea only and does not discuss GR00T 2.
The key insight: As we read Nvidia’s research page, the scarce input for GR00T 2 was not robot time: human video is 92.3% of its 436,414 pretraining hours by our arithmetic, and five demonstrations per task take it from 40% to 95% success on 50 new tasks, by Nvidia’s count.
The research page describes GR00T 2 as a model that “predicts future video and robot actions together”, and says it runs on parallel grippers, dexterous hands and humanoids. 100,000 hours of the human video are “openly available as Human-1M-100K on Hugging Face”; the dataset page lists a CC BY 4.0 licence.
On 50 tasks the target robot had never seen in its own data, the page says GR00T 2 “succeeds 40% of the time zero-shot”. One demonstration per task lifts success to 71%, and five lift it to 95%, “after a 30-minute fine-tune on a single RTX 5090”.
The blog says GR00T 2 “currently ranks among the top three models across all vision language action and world action model categories for the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks”. Those rankings are Nvidia’s own claim.
Hours of pretraining data for Nvidia’s GR00T 2, by kind, from Nvidia’s GEAR research page on GR00T 2. Shares of the 436,414-hour total are our arithmetic.
How It Compares With GR00T 1.7
The research page gives one direct comparison with the previous generation. With simulated data added, “GR00T 2 Turbo’s out-of-distribution success rises from 6.7% to 63.3%. GR00T 1.7 rises to 40.0%”. GR00T 2 Turbo is a smaller 2B-parameter version; the page says a prediction takes 124 ms on the Jetson Thor computer inside a Unitree G1 humanoid.
For scale on the earlier versions, Hugging Face’s API showed GR00T-N1.7-3B at 454,340 all-time downloads and GR00T-N1.6-3B at 282,574 when we checked on 9 November. Our earlier piece covered GR00T N1.6’s download count.
Who Is Using It
Nvidia’s blog says “Nexuni is testing it for robotic laundry handling, Foxconn is applying GR00T 2 to factory automation for server manufacturing and Siemens is evaluating it across industrial use cases”. It adds that MicroAGI uses GR00T 2, post-trained with reinforcement learning, to swap Ethernet cables with a YAM robotic arm.
The same post describes Isaac Lab 3.0, the latest release of Nvidia’s open-source robot-learning framework, and says it reached “up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX” on a dexterous manipulation benchmark. For a humanoid built on the earlier model, see Agile Robots’ reference humanoid on GR00T, and for the wider field, 87 humanoid robot players, from Nvidia to CATL.
The Structural Read
As we read the two Nvidia posts, the model, the 100,000-hour human-video dataset, the Isaac Lab framework and the benchmarks are all released as open tools, while the stated runtimes are Nvidia hardware: a single RTX 5090 for the 30-minute fine-tune and Jetson Thor on the humanoid.
As we read the partner list, the named uses are industrial: Foxconn on server manufacturing, Siemens across industrial use cases, Nexuni on laundry handling and MicroAGI on Ethernet cable swapping, all as described by Nvidia.
NVIDIA GEAR research page, GR00T 2
“Learning to predict the world, not just act in it, unlocks three new capabilities.”
Three Implications
FOR ROBOT MAKERS Nvidia’s page says five demonstrations per task, under 3 minutes of robot data, reach 95% success after a 30-minute fine-tune on a single RTX 5090.
FOR DATA BUILDERS Nvidia says 100,000 hours of its human video are openly available as Human-1M-100K on Hugging Face, under a CC BY 4.0 licence per the dataset page.
FOR WATCHERS OF THE RELEASE We found no GR00T 2 checkpoint among Nvidia’s Hugging Face models on 9 November; GR00T-N1.7-3B showed 454,340 all-time downloads.
The framework says: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Nvidia opens the model, the dataset and the training framework, while the fine-tuning and on-robot runtimes it names are its own chips. The framework predates GR00T 2 and does not discuss it.
What Is Not Established
Nvidia calls GR00T 2 open, but the announcement gives no release date for model weights. In four searches of Nvidia’s Hugging Face models on 9 November we found GR00T N1 to N1.7 checkpoints and none named GR00T 2. The benchmark rankings and task success rates are Nvidia’s own figures, and the shares of the training mix are our arithmetic.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Nvidia’s GR00T 2 is pretrained on 436,414 hours of data, 92.3% of it human video by our arithmetic, and reaches 95% success on 50 new tasks after five demonstrations each, according to Nvidia’s research page.
A note on sourcing. Training-data hours, task success rates, benchmark rankings, partner uses and quotations are from Nvidia’s 9 November 2026 CoRL blog post and its GEAR research page for GR00T 2. Download counts and the dataset licence are from Hugging Face, read on 9 November. Shares of the training mix are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
