Robotics stopped being about the hand a long time ago. It's about the head. What's new in this model is that the head watches the footage of its own work.
Fact
The facts: On July 30, 2026, Google unveiled Gemini Robotics ER 2, described by the company as a high-level brain for robots. The model processes a video stream, tracks how far a task has progressed, and corrects itself during execution. By its own measurements, progress-classification accuracy is 57.4%, and finding the right frame for a key event is 91.3% accurate, with 0.96 seconds of latency. The model coordinates tools and action models across multi-step tasks and lets different robots work together. Among the described capabilities is recognizing a nearby person and stopping work. It's available through the Gemini API and the Google AI Studio environment; the enterprise platform is in closed preview. All the figures come from Google. Source: Google, 30.07.2026.
Hold the enthusiasm for a second. The number everyone will quote is 91.3 percent - the accuracy at spotting the key frame. The number that says more is the other one: under 60 percent accuracy at judging how far a task has gotten. On a shop floor that means on almost every other attempt the machine isn't sure it's done. Great for a demo. Not for a shop floor.
A robot that knows it got something wrong is more useful than a robot that doesn't get anything wrong in the clip.
The part that genuinely pleases me is stopping when it spots a person nearby. It's not flashy and won't make a headline, but it's exactly what decides whether a robot like this can stand in a room with people. Everything else is performance. This is a permit.
This is a good step toward robots that check themselves, and it's presented honestly - the numbers are published, including the weak one. The real test, though, is on a floor that's greasy, under bad light, with nobody clearing objects out of the way.
Official primary sources
Google - Gemini Robotics ER 2, 30.07.2026The visual is generated code-art, no third-party images.