Robots That Work Offline – Google Unveils Revolutionary AI Model That Does Not Require the Internet

Imagine a robot that understands you, can manipulate objects, and adapt to your verbal commands—all without an internet connection. Sounds like science fiction? Think again! This is exactly what Google DeepMind has now achieved with its new version of the Gemini Robotics AI model.

Gemini Robotics On-Device: An AI Model That Does Not Require a Cloud Connection

The new version of the Gemini Robotics model works on the same principle as the previous VLA (vision-language-action) model. This vision-language-understanding principle ensures that the robot not only sees what you are doing and understands what you are telling it, but can also act accordingly. However, the new Gemini Robotics On-Device version is also optimized to run directly on robotic devices without requiring an internet connection.

While remote servers handled command processing until now, it is now done directly on the robot's hardware. This enables fast responses even in places with unstable or no internet connection, as well as a higher level of privacy and security.

Improved Capabilities and Rapid Learning

According to benchmark tests, the previous cloud-based model is still more powerful, but the results are very close. This indicates that the offline capabilities are already at a very high level.

Gemini Robotics On-Device is revolutionary not only because it runs directly on robots, but also because it can learn very quickly. It can generalize to new situations, understand new things, and respond to them immediately.

Improved Response to Commands

A robot equipped with the Gemini Robotics On-Device AI model can perform pre-learned tasks from the outset. However, it can also adjust to tasks for which it was not primarily trained and adapt to new instructions.

It needs only about 50-100 demonstrations to understand and correctly perform a new task. Gemini Robotics On-Device is therefore very useful in situations that require quick responses to commands and adaptation to new circumstances. All without an internet connection.

Trained on Robots

Gemini Robotics On-Device is trained on the ALOHA robot, but it can also adapt to other models. Its capabilities were also confirmed during training on Apptronik's Apollo humanoid robot and the dual-arm Franka FR3 robot.

According to the available demonstrations, robots equipped with this on-device model can handle a wide range of commands involving fine motor skills, such as:

• folding clothes,

• placing fruit in a bowl,

• cleaning a work surface,

• unzipping a zipper

• or putting items into a bag.

Development and Testing

Google DeepMind is releasing an SDK and allowing the model to be adapted to specific robots—but for now, only to its selected partners and carefully chosen, trusted testers. The aim of this step is to gradually verify its safety and applications before making it available to the wider public.

Gemini Robotics On-Device represents a significant step toward robot autonomy. This new AI model can run offline inside a robot while still adapting flexibly to new tasks.

Source: Google DeepMind, The Verge, TechCrunch