Google DeepMind presents Gemini Robotics On-Device, an AI model that works without the Internet

Google DeepMind, a team of artificial intelligence researchers, has presented its new AI model for robots called Gemini Robotics On-Device. They say that it is their most powerful VLA model that can work locally without the Internet.

It is based on the capabilities of the previous Gemini Robotics model, which was released in March. Gemini Robotics On-Device was created for “two-armed” robots with minimal computing resources.

The blog says that the new model understands instructions in natural language, performs very precise and complex tasks (for example, unzipping a bag or folding clothes). And all of this is done directly on the robot itself, without the Internet or cloud processing.

In benchmark comparisons, Google Gemini Robotics On-Device performs almost on par with cloud-based Gemini Robotics. The company claims that it outperforms other on-premises models in common benchmarks, but did not specify which ones were used for comparison.

During the demonstration, Google showed how robots running on the local Gemini Robotics On-Device model perform tasks such as unzipping bags and folding clothes.

The model was initially trained for ALOHA robots, but later adapted to work on the Franka FR3 bipedal robot and the Apollo humanoid from Apptronik.

The researchers say that Franka FR3 successfully coped with new tasks and objects that the model had never seen before, such as assembling parts on a production line.

Google DeepMind also releases Gemini Robotics SDK, a tool for developers. It allows training robots to perform new tasks by showing them only 50–100 demonstrations.