What’s the story
Google DeepMind has announced an on-device version of its Gemini Robotics AI model.
The new vision-language-action (VLA) model is small and efficient enough to run directly on a robot, without needing an internet connection.
This marks a major step forward in robotics technology. The on-device AI model can help robots perform a variety of physical tasks even if they haven’t been specifically trained for them.
Offline model nearly as good as flagship Gemini Robotics version
Carolina Parada, head of robotics at Google DeepMind, explained that the original Gemini Robotics model uses a hybrid approach. This allows it to operate both on-device and in the cloud.
However, with this new device-only model, users can access offline features that are almost at par with those of the flagship version.
The on-device model can perform several different tasks out of the box and adapt to new situations with just 50-100 demonstrations.
On-device model trained only on Google’s ALOHA robot
Google trained the on-device model only on its ALOHA robot but was able to adapt it to different types of robots. These include Apptronik’s humanoid Apollo robot and bi-arm Franka FR3 robot.
Parada said, “The Gemini Robotics hybrid model is still more powerful, but we’re actually quite surprised at how strong this on-device model is.”
Along with the launch of this innovative tech, Google is also releasing a software development kit (SDK) for developers to evaluate and fine-tune it.
