Nvidia’s GR00T 2 Lets Robots Master Delicate Tasks by Watching Human Video

Nvidia has introduced GR00T 2, an open AI model designed to help robots acquire delicate manipulation skills by observing human video, and it is releasing a massive 100,000-hour first-person video dataset for anyone to use. Announced at the Conference on Robot Learning (CoRL) in Austin, Texas, on November 9, 2026, the model is described as a “world action model” that predicts how its movements will alter its visual surroundings before executing them.

• Open model: GR00T 2 is an open world-action model trained on robot data and human video, including 100,000 hours of first-person footage released as Human-1M-100K.
• Strong benchmarks: It completed 40 of 50 tasks without task-specific training, achieving 71% average success with one demonstration and 95% with five.
• Industrial trials: Foxconn, Siemens, and Nexuni are testing GR00T 2 for server manufacturing, factory work, and laundry handling, respectively, though none has confirmed production use.

How GR00T 2 Learns and Transfers Skills

At its core, GR00T 2 combines robot data with human video to transfer skills from people to machines. According to Nvidia, the model was pretrained on a mixture of 402,730 hours of egocentric human video, 10,184 hours of handheld-gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation. This allows robots to learn from human demonstrations and pass skills between robots with different physical bodies without additional robot demonstrations.

The release of the 100,000-hour Human-1M-100K dataset is arguably the most significant move, as the bottleneck in robotics is not hardware but training data. Nvidia previously showcased a similar approach with DreamDojo, a “world model” that taught robots by watching 44,000 hours of human video, and GR00T 2 more than doubles that effort. The company also released Isaac Lab 3.0, an open-source tool for training robots in simulation, which it says trains up to 2.5 times faster on one benchmark compared to Isaac Lab 2.3 with PhysX.

GR00T 2 uses a 2-billion-parameter transformer built on Cosmos-Predict2.5-2B, combining recent camera views, robot state, and a written instruction to generate expected future views and the next action sequence. One pretrained checkpoint covers parallel-gripper arms, dexterous hands, and a humanoid. Cross-embodiment tests show skills transferring between bodies: a robot with two five-fingered hands lined a lab funnel with filter paper and wetted it with a wash bottle in 65% of trials after being shown only by humans, and a YAM arm assembled a toy dump truck end-to-end in 60% of trials after learning from a handheld UMI gripper.

Benchmarks and the Fine Print

Nvidia’s own figures are impressive: GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations, achieved 71% average success with just one demonstration per task, and reached 95% with five demonstrations. Those five demonstrations supply under three minutes of robot data per task, followed by an approximately 30-minute adapter fine-tune on a single RTX 5090.

However, Nvidia’s technical whitepaper acknowledges a caveat: the evaluated tasks were selected among those the model can learn with five episodes at a success rate of at least 80%, so the average is high partly by construction. That does not invalidate the results, but it means they should be read as a showcase rather than a real-world average. The model currently ranks among the top three across all vision-language-action and world-action model categories for benchmarks including RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces, with multiple leaderboard wins.

“To me, agentic robotics is the most exciting technological development since the internet in 1995,” said Ken Goldberg, a robotics professor at UC Berkeley who worked with Nvidia on two of the projects.

Industrial Trials and the Road Ahead

Foxconn, which now earns most of its revenue from AI servers, is using GR00T 2 to help build them. Siemens is testing it for factory work, and Nexuni is trying it for laundry handling. None has confirmed real production use yet. Siemens has separately tested an Nvidia-powered humanoid on an eight-hour factory shift, though that trial was not specifically for GR00T 2. Other partners include MicroAGI, which is using GR00T 2 post-trained through reinforcement learning to perform Ethernet cable swapping with a YAM robotic arm.

Nvidia also introduced Isaac Lab-Arena, an open-source framework for scaled policy evaluations, and DexBench v1.0, developed with RLWRLD, containing 15 tasks across retail, logistics, assembly-line, and hotel food-and-beverage scenes. The company is also working with the University of Washington on Success-Guided Sampling, which directs practice toward configurations near the policy’s capability frontier and has shown effective scaling beyond one million simulated environments.

My take is that GR00T 2 matters more for what it distributes than for what it boasts. The 100,000 hours of released video are worth more than any benchmark because they attack the real problem—the lack of data—and do so openly, so that not only Nvidia can build on them. That said, enthusiasm should be tempered: a model that hits 95% on tasks chosen by its creator, and three companies that “test” it without serious deployment, is exactly where humanoid robotics has been for years—the constant promise that never quite lands on the factory floor. Goldberg’s “internet in 1995” line is seductive, but in 1995 the internet already worked for millions of people; these robots, for now, work for demo videos. The progress is real and the direction is right. What is missing is the usual: getting out of the lab.