Nvidia's GR00T 2 hits 95% task success after five demos

Nvidia reports that GR00T 2, its new robot foundation model, reached 95% average success on 50 unseen tasks after just five demonstrations each — under three minutes of robot data — followed by a 30-minute fine-tune on a single RTX 5090. The company launched the model together with version 3.0 of its Isaac Lab training framework on Monday at the Conference on Robot Learning (CoRL) in Austin, Texas.

GR00T 2 is what Nvidia calls an "open frontier world action model." It succeeds the GR00T N1 series and is built on Nvidia's Cosmos world models. The core idea is simple to state: the model predicts what the robot's camera will see next and what the robot should do, in a single step.

How does the model work?

A world action model generates future video and robot actions together. GR00T 2 reads recent camera frames, the instruction and the robot's current state, then produces a short chunk of predicted future video plus motor commands inside one diffusion transformer, with small output heads tuned per robot, according to Nvidia's research page.

A faster variant, GR00T 2 Turbo, uses a 2-billion-parameter transformer and completes a prediction in 124 milliseconds on Nvidia's Jetson Thor robot computer. The takeaway: the architecture ties perception and control into one network that runs fast enough for on-board robot hardware.

What went into the training data?

The pretraining mix is dominated by human video:

• 402,730 hours of first-person human video
• 10,184 hours of handheld gripper demonstrations
• 23,500 hours of robot teleoperation across arms, dexterous hands and Unitree's G1 humanoid

Nvidia released 100,000 hours of the human video as an open dataset on Hugging Face. Its researchers report that adding all of the human video raised real-robot task progress from 50% to 70% across 72 tasks. The plain reading is that watching people do things transfers directly to robot performance.

How well does it perform?

On 50 tasks the model had never seen, GR00T 2 averaged 40% success with no demonstrations at all, and completed 40 of the 50 at least once. One demonstration per task lifted the average to 71%; five demonstrations brought it to 95%.

Nvidia also says the model ranks in the top three across vision-language-action and world action model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks. All of these results are Nvidia's own, not independently verified. The research page lists the code as "Available in December" and states no license or release date for the model weights, so outside validation is not yet possible.

What is new in Isaac Lab 3.0?

Isaac Lab 3.0 lets AI agents set up training environments, switch between the PhysX and Newton physics engines, and apply a new reinforcement-learning post-training library to models such as GR00T. Newton adds the ability to simulate cables, cloth and multi-joint robot hands. Nvidia says the update reaches up to 2.5 times the training throughput of Isaac Lab 2.3 on a dexterous manipulation benchmark.

The company also shipped Isaac Lab-Arena 0.4 as a developer preview for evaluating policies across thousands of environment variations in parallel. In short, the simulation side of Nvidia's robotics stack is now both faster and broader in what it can model.

Who is using it already?

Early adopters span factory floors and research labs:

• Foxconn is applying GR00T 2 to server manufacturing
• Siemens is evaluating it for industrial work
• Nexuni is testing it for robotic laundry
• MicroAGI post-trained the model to swap Ethernet cables on a robot arm
• Skild AI announced STEEL, a rigid-body physics solver built on Newton

The release follows Nvidia's recent opening of four Cosmos 3 simulation models and its humanoid reference systems. With code promised for December and the weights' licensing still unstated, the next signal to watch is whether independent labs can reproduce the reported success rates once the model leaves Nvidia's own benchmarks.