Nvidia unveiled GR00T 2, an open AI model meant to teach robots delicate hand skills, trained partly on 100,000 hours of human video.
In Nvidia's own tests, the model reached 95% average success with five demonstrations per task. A demonstration is a worked example of the job, and five is a small number. Foxconn, Siemens and Nexuni are testing the model. The available reporting does not say which tasks were used or what earlier models scored, so the 95% has no baseline to be measured against yet. It also calls the model "open" without spelling out what that covers.
The design choice is the interesting part. Robotics has a data problem that chatbots do not. Language models learned from a huge archive of text that already existed. Nothing comparable exists for robot hand movements, and collecting it means someone physically guiding machines through every task. Video of people doing things with their hands is far easier to find. Training on 100,000 hours of it is a bet that human footage can stand in for some of that missing robot data.
The five-demonstration figure matters for the same reason. For a factory customer, the cost of a new robot skill is mostly the effort of teaching it. If five examples really are enough, that cost falls sharply.
Two cautions apply. Testing is not deployment, and the three companies named are evaluating the model, not running it in production. And 95% average success means something different on a production line, where a failure can stall the work around it, than in a lab trial. The missing task list is what would settle which of those readings fits.