Nvidia's new GR00T 2 teaches robots by watching humans

Nvidia pulled back the curtain Monday on GR00T 2, a humanoid robotics AI model built to teach robots tricky physical work — grabbing, moving and manipulating objects — by copying what humans do. The pitch is simple: show the robot, don't program it. Nvidia says the model also carries those learned skills over to different robot bodies, so a trick learned on one machine isn't stuck there.

The model is already out in the wild. Nexuni is running it on robotic laundry handling, and Foxconn is using it to automate parts of server manufacturing, according to Breaking The News. Siemens is evaluating GR00T 2 across a spread of industrial applications while it weighs a wider rollout.

Nvidia is pairing the model with an open-source robot-learning framework aimed at researchers. The company says agents can quickly spin up training environments using simulation-ready assets, pick renderers for perception-based training, and bolt a new reinforcement post-training library onto foundation models like GR00T.

That combination — a model that transfers across hardware plus open tooling for the research crowd — puts Nvidia in the middle of the humanoid race on both the chip side and the software side. For now, the proof is in the factories: laundry, servers and Siemens' test benches.

headline: Nvidia's new GR00T 2 teaches robots by watching humans

entities: [