NVIDIA unveils open GR00T 2 robotics model
NVIDIA has released GR00T 2, an open "System One" robotics model that learns from human demonstrations and transfers skills across robot bodies, so each robot need not learn every task from scratch.
NVIDIA announced GR00T 2 on X, calling it its "most advanced, open System One model for robotics research."
The company says GR00T 2 learns skills from human demonstrations. It also predicts how its own actions will change the physical world. NVIDIA describes this as a model of the consequences of physical actions, not only a mapping from observations to motor commands.
NVIDIA's main claim is cross-embodiment transfer. Skills learned by the model "can carry across different robot bodies," which the company says reduces the need to teach each robot every task from scratch. If this holds up, it would lower the cost of bringing a new robot platform up to speed. Today, collecting data for each platform separately is a major bottleneck.
The post includes a video thread showing the model running on hardware. NVIDIA describes the demos as "real robots, real time" and "fully autonomous."
The announcement is light on technical detail. It does not give model size, architecture, training data, benchmark results, or comparisons with the earlier GR00T models. It also doesn't say which robot bodies were tested, what license "open" means here, or where the weights and code can be downloaded. In NVIDIA's terminology, "System One" usually means the fast, reactive control layer, not slower deliberative planning. The post doesn't say how GR00T 2 connects to a higher-level reasoning component.
Researchers will need a technical report or model card to judge how well the transfer claims hold up.