DeepMind’s Gemini robotics now powers robots offline, raising concerns over autonomous AI

Google DeepMind has unleashed a game-changing update: its Gemini Robotics model can now run entirely offline, letting robots function autonomously without cloud support. The streamlined vision-language-action (VLA) system offers nearly the same dexterity as the cloud-connected flagship model—but is compact enough to reside wholly on-device.

This fresh development empowers robots to see, understand spoken instructions, and act with fine motor skills—all while disconnected from the internet. Previously, engineers relied on a hybrid system combining on-board intelligence with cloud computation. Now, developers can deploy Gemini Robotics in remote or secure environments where connectivity is unreliable or restricted.

Carolina Parada, DeepMind’s Head of Robotics, describes the device-only model as “a starter model” that impressively delivers performance comparable to the full hybrid version. She expects it to be ideal for sensitive applications or locations with low bandwidth.

Importantly, the offline AI retains Gemini’s ability to generalise to new tasks with minimal input—only around 50 to 100 examples are needed for it to learn something new. Initially trained on DeepMind’s ALOHA robot, this version has already been tested on the Apptronik Apollo humanoid and the dual-arm Franka FR3, demonstrating notable cross-platform adaptability.

To support broader experimentation, DeepMind is launching a development kit (SDK) for the on-device model. For the first time, external teams can evaluate and fine-tune a Gemini VLA directly—shifting the model from internal R&D into the hands of trusted robotics developers.

The company stresses the model and SDK are currently restricted to a carefully selected group of testers while they assess safety and ensure its responsible deployment.