A robot can follow spoken instructions, unzip a bag and fold clothes while its AI model runs locally. Google DeepMind described those demonstrations when it introduced Gemini Robotics On-Device in June 2025.
The distinction is where decisions happen. A model running on the robot can continue operating without a network connection. Google presents that as useful when delays matter or connectivity is unreliable.
This is a vision-language-action model: it connects camera input and language instructions with movements. Understanding a request is only one part of the job. The machine also has to manipulate a physical object.
Google says the model adapted to new tasks with as few as 50 to 100 demonstrations. That number describes its reported experiments, not a guarantee that any robot can learn any task from that amount of data.
The team trained the system on ALOHA robots and adapted it to other hardware, including a two-arm Franka setup and Apptronik's Apollo humanoid. Moving between machines matters because their bodies and control systems differ.
The company's own comparisons also showed a limit. The larger Gemini Robotics system performed better on more demanding tasks without the same on-device constraints. Local operation comes with a capability trade-off.
The initial release targeted a trusted tester group. Developers still needed to evaluate the complete physical system, including its safety controls. The demonstrated robot could work without a network; that did not remove the need for testing.
