Gemini Robotics On-Device

Gemini Robotics On-Device, announced by Google DeepMind in June 2025, is a vision-language-action model built to run entirely on a robot's local compute with no cloud connection. It brings the Gemini Robotics foundation model's dexterous manipulation capabilities, including bi-arm control on platforms like ALOHA, the Franka FR3 and the Apptronik Apollo, to settings where latency, privacy or connectivity rule out cloud inference.

Google reports that the model can adapt to new tasks with only 50 to 100 demonstrations, and it ships alongside a Gemini Robotics SDK. Access is not fully open: availability runs through a trusted tester program, and weights are closed. It is nonetheless a landmark release because it moves capable generalist robot policies from the data center onto the machine itself.