NVIDIA has taught a robot model to learn by watching people. GR00T 2, introduced on November 9 at the Conference on Robot Learning in Austin, is an open world action model built on NVIDIA Cosmos that picks up dexterous manipulation skills from human video. It completed 40 of 50 test tasks at least once with no task-specific demonstrations, reached 71 percent success with a single demonstration per task, and 95 percent with five. NVIDIA is also opening 100,000 hours of the human video it trained on.
One model, many bodies
GR00T 2 is a two-billion-parameter transformer built on Cosmos-Predict2.5-2B, according to NVIDIA’s research page. It watches recent camera views, the robot’s state and a written instruction, then predicts the future views and the next sequence of actions. A single pretrained checkpoint covers parallel-gripper arms, dexterous hands and a humanoid body, so the same model can drive very different machines.
The training data is the point. Pretraining combined 402,730 hours of first-person human video with 10,184 hours of handheld-gripper demonstrations across 2,897 tasks and 23,500 hours of robot teleoperation. The Human-1M-100K release opens 100,000 of those human hours, with video, language annotations and head-and-hand motion data, under a CC BY 4.0 licence that allows commercial use.
What the numbers actually say
NVIDIA reports the model completed at least one trial on 40 of the 50 tasks without task-specific demonstrations. With one demonstration per task, success reached 71 percent, and with five it reached 95 percent. Those five demonstrations add up to under three minutes of robot data per task, followed by about a 30-minute adapter fine-tune on a single RTX 5090.
Skills move between bodies. NVIDIA reports 65 percent success on a filter-paper setup learned from human demonstrations with no task-specific teleoperation, and 60 percent on a toy dump-truck assembly transferred from a handheld gripper to a YAM arm. The model ranks among the top three across vision-language-action and world action model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks.
Simulation does the rehearsing
The other half of the announcement is the training ground. Isaac Lab 3.0, NVIDIA’s open-source robot-learning framework, is generally available since October 20 and now lets agents set up training environments quickly, switch between PhysX and Newton physics solvers, and apply a new reinforcement post-training library 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 Isaac Lab 2.3 with PhysX. Two new open benchmarks arrive with it: DexBench, built by RLWRLD with NVIDIA, tests dexterity across 15 tasks in retail, logistics, manufacturing and hospitality scenes, and VABAR, from UC Berkeley and NVIDIA, covers six industrial tasks.
Agents get a bigger role too. UC Berkeley’s Graph-as-Policy framework uses agents to connect individual skills into robot programs, rehearsing them in simulation and revising them after failures. NVIDIA says a reasoning agent sequences GR00T 2’s skills, monitors progress and invokes recovery routines when parts drop or work stalls.
Who is already trying it
Nexuni is testing GR00T 2 for robotic laundry handling, Foxconn is applying it to factory automation for server manufacturing, and Siemens is evaluating it across industrial use cases. MicroAGI is using the model, post-trained through reinforcement learning, to perform Ethernet cable swapping with a YAM robotic arm. None of the companies has announced a production rollout.
Our Read
The headline numbers deserve a careful read. Completing a task at least once is not a success rate, and NVIDIA’s own white paper says the five-demonstration tasks were selected from ones the model can already learn at 80 percent success or better, so the 95 percent figure is high partly by construction.
The honest test is the 71 percent one-demo average, and even that comes with an agent that waits during a median 38 percent of an episode in one sorting test. The model is also described as open, but its code is due in December, so researchers cannot yet check the claims themselves.
The direction matters. Robots that learn from watching people need far less robot-specific training data, and the open dataset lowers the barrier for every lab working on the same problem. Humanoid robotics had a busy week: Agility Robotics went public on Nasdaq earlier, and NVIDIA’s bet is that the winning robots will be the ones that rehearse in simulation first.
Why It Matters
Dexterous manipulation is the bottleneck between robot demos and useful factory workers. A model that generalises from human video to new tasks, and transfers skills between different robot bodies, attacks that bottleneck. If the open release in December holds up, GR00T 2 could become the starting point for humanoid work, the way open language models became the base for chatbots.
The test that matters now is deployment. Foxconn and Siemens are evaluating, not buying, and the gap between a lab benchmark and a factory floor is where most robot models stall. Watch whether the December code release brings independent replications, and whether any of the early testers moves from trials to a production line.
Background
GR00T 2 builds on NVIDIA’s Cosmos 3 open world models, released days earlier for cars and robots, and follows the physical-AI work NVIDIA showed at its Seoul showcase, where an LG robot running GR00T demonstrated factory tasks. The model arrives as NVIDIA’s robotics stack, from simulation to benchmarks, becomes the company’s second front after AI data centres.
Source: NVIDIA’s announcement and Unite.AI’s report
0 Comments