studio raoulsoftware innovation, architecture & development
← Blog
AI in Practice · November 10, 2026

Robots that learn by watching — and what that means for factories

NVIDIA's new GR00T 2 model lets a robot master a task after seeing it demonstrated just a few times, making it far easier for real businesses to put robots to work.

Key facts
95%
success rate after 5 demonstrations
71%
success rate after just 1 demonstration
100,000 hrs
of human video used to train the model
50
different tasks tested

Show a robot something five times, and it can do it reliably. That is what NVIDIA's new GR00T 2 model — released Monday at a robotics conference in Austin, Texas — achieved in controlled tests: a 95% success rate across 50 tasks after just five demonstrations each. Even after a single demonstration, it succeeded 71% of the time.

Until now, training a robot to do something new meant weeks of custom programming and expensive specialists. GR00T 2 takes a different approach: it was trained on 100,000 hours of first-person human video, so it already understands a wide range of physical movements before it ever arrives in a factory. That shared foundation means the same model can work across very different robots and very different tasks.

Foxconn is already applying it to automate parts of its server assembly process in Houston. Siemens is evaluating it for factory use. These are not small experiments — Foxconn builds some of the most precisely manufactured hardware on the planet, and real factories are already betting on this. NVIDIA's CEO Jensen Huang has said every industrial company will become a robotics company.

If your business involves physical, repetitive work — assembly, packaging, picking, sorting — this is worth watching closely. The barrier to deploying a robot has dropped: you no longer need a team of engineers to program it task by task. You show it the work, and it learns. Keep in mind these are controlled test results; real environments are messier. The practical step today: pick the first repetitive task in your operation that a robot could plausibly handle, and start paying attention to how early adopters like Foxconn actually perform.

Sources
← Agent work just got 17x cheaper — what that changes for builders← The AI tools in your stack just became permanent infrastructure
Robots that learn by watching — and what that means for factories