With Google's Gemini Robotics On-Device paradigm, voice commands may be followed offline by robots, allowing for quick, all-purpose activities without internet access.

A Gemini model developed by Google DeepMind can operate robots without an internet connection. According to Google, the Gemini Robotics On-Device model, a vision-language-action (VLA) system, has been fine-tuned to run directly on the robot, without relying on external computing.

The company states that the new VLA model, developed from the cloud-based Gemini Robotics, demonstrates robust general-purpose skills and the ability to perform a wide range of tasks. Gemini Robotics On-Device can be useful for tasks requiring prompt responses to commands or in settings with spotty or nonexistent internet connectivity because it can be utilized to carry out tasks without a data network.

The model is designed for bi-arm robots that require less processing power to operate. Google demonstrated a robot running the model locally and responding to voice commands to open a container, unzip a bag, and uncap a marker.

The VLA model was initially trained exclusively for ALOHA robotics, but Google claims to have successfully adapted it for use with Apptronik's Apollo humanoid system and the dual-arm Franka FR3 robot. While operating the model, the Franka robot was able to carry out activities like replacing industrial belts and folding garments.

Google said that it was able to modify the model of the Apollo humanoid robot to follow natural language commands and generally manipulate various things, including ones that had never been seen before. Through its trusted tester program, Google is making the Gemini Robotics software development kit (SDK) available to developers so they can quickly assess Gemini Robotics On-Device in their settings and workloads.