Google DeepMind has unveiled a powerful new version of its Gemini Robotics AI that runs directly on robots without an internet connection. Designed to perform complex tasks in real-time, this smaller, more efficient model brings impressive intelligence and adaptability to your machines.
Gemini Robotics AI now works without internet access
If you work with robotics, you know that most AI-powered systems rely on cloud connectivity. But now, Google DeepMind is changing that with its on-device version of Gemini Robotics. This updated version of their vision-language-action (VLA) model doesn’t need Wi-Fi or mobile data — it can run fully offline.
This means your robot can “see,” “understand,” and “act” all on its own. It can perform physical tasks and even adapt to new challenges, just like the flagship Gemini Robotics model first released in March. But the key difference? This one is compact and efficient enough to run directly on the robot without external processing.
According to Carolina Parada, head of robotics at Google DeepMind, the original Gemini Robotics model worked using a hybrid system — part cloud, part device. This new version, however, delivers most of its power entirely offline, making it ideal for environments with weak internet or high privacy needs.
Fast learning and adaptability built-in
What’s especially impressive about this on-device AI model is how quickly it can learn. You only need 50 to 100 demonstrations for it to pick up a new task. If you want your robot to start sorting packages, setting a table, or handling delicate objects, it can get up to speed with just a few examples.
Although the model was trained on Google’s ALOHA robot, its versatility has already been proven. It has been adapted for other robots, including the humanoid Apollo robot by Apptronik and the bi-arm Franka FR3 robot.
Parada admits the hybrid version of Gemini Robotics still outperforms this offline model in some areas. However, she says the standalone version is surprisingly powerful and could be ideal as a “starter model” or in industries prioritising data security.
Developers can now get early access to fine-tune the AI
To support this rollout, Google DeepMind is also launching a software development kit (SDK) for the on-device model. This SDK will allow developers like you to test the AI, make adjustments, and better tailor its performance to your robot’s needs. It’s the first time the company has offered such a toolkit for one of its vision-language-action systems.
While the on-device model and its SDK will not be widely available, a select group of trusted testers will get early access. Meanwhile, DeepMind will ensure the technology meets safety standards before a public launch.
If you’re developing robotics solutions and need high-performance AI in offline or secure settings, this latest innovation from Google DeepMind could be just what you’ve been waiting for.
