Teaching a robot a new job now takes hours, not months
Google's Gemini Robotics 2 can adapt to a brand-new robot body in a few hours — which changes the biggest bottleneck in commercial robotics.
- 92%
- success rate unscrewing a light bulb
- < 200 examples
- data to adapt to a new robot body
- 3
- AI models working together
- Jul 30, 2026
- release date
Building a useful robot has always needed two expensive parts: the physical machine and the software to control it. The software had to be written almost from scratch for every new robot — a process that took months. Google DeepMind released Gemini Robotics 2 on July 30, and it changes that fundamentally. The system is a single AI that runs any robot body — and learns a new one in just a few hours.
The system is actually three models working together. One does high-level planning, deciding what the robot should do next. Another translates those plans into precise movements, keeping the robot balanced from head to toe. A third version runs entirely on the robot's own computer, no internet required. If something goes wrong mid-task, the robot finds the last step that worked and tries again from there, on its own.
The results so far are uneven but real. Unscrewing a light bulb — a task that requires fingers, wrist, elbow, and full-body balance all working together — succeeded with a 92% success rate. Tying a trash bag worked 44% of the time. Not perfect, but a clear jump from where this technology was just a year ago.
The biggest change for anyone buying or building robotic systems is how fast the AI can transfer to new hardware. Training takes just hours and fewer than 200 examples, not months of custom engineering. Google provides the AI brain; hardware partners like Boston Dynamics and Apptronik supply the robot bodies. That split matters: bring in new hardware and have it running the same day. If you are evaluating robotic systems right now, ask your vendor how long retraining takes for a new machine — that number just became the most important one.