Google DeepMind has launched Gemini Robotics On-Device, a compact, efficient AI model that allows robots to operate entirely offline, without relying on cloud connectivity. This is a major leap for the Gemini Robotics project, which was previously dependent on internet access for decision-making and task execution.
According to Carolina Parada, head of robotics at DeepMind, the new version is “small and efficient enough to run directly on a robot,” making it ideal for real-world environments like remote areas, secure facilities, and latency-sensitive scenarios. She described it as a “starter model” suitable for applications with poor or no internet access.
Despite its smaller footprint, Gemini On-Device has shown surprisingly strong capabilities. It can:
-
Perform tasks out of the box
-
Learn new tasks from just 50 to 100 demonstrations
-
Execute detailed actions like folding clothes or unzipping bags
-
Function on various robots such as Google’s ALOHA, Apptronik’s Apollo humanoid, and Franka FR3
Unlike Tesla’s Optimus robot, which needs to send data to cloud servers, Gemini On-Device processes all information locally. This makes it especially suitable for privacy-sensitive sectors, such as healthcare and industrial automation.
While the full cloud-based Gemini Robotics model includes semantic safety tools, the on-device version does not. Developers are encouraged to use Gemini Live APIs and integrate external safety mechanisms.
This launch aligns with Google's broader push for offline AI. It follows the release of AI Edge Gallery, an Android app that enables users to run AI models offline on smartphones using the compact Gemma 3 1B model and frameworks like TensorFlow Lite.
Together, these innovations signal Google’s emphasis on private, low-latency, and autonomous AI systems – both for mobile devices and next-generation robotics.
