Google DeepMind Unveils On-Device Gemini AI for Robots

Google DeepMind is introducing a compact version of its Gemini Robotics AI model that can run directly on robots without requiring internet access. This new on-device vision-language-action (VLA) model mirrors the fine motor capabilities of its cloud-powered predecessor, but with the added benefit of full offline functionality.

Designed to help robots perform complex tasks—even ones they haven’t been explicitly trained on—this model enables generalization, command interpretation, and adaptive manipulation in various physical environments.

“It’s small and efficient enough to run directly on a robot.” — Google

The on-device variant still draws from the hybrid principles of the original Gemini Robotics system, which can operate both in the cloud and locally. However, this version focuses solely on edge-device performance, making it ideal for deployments in areas with poor connectivity or strict data security requirements.

“The Gemini Robotics hybrid model is still more powerful, but we’re actually quite surprised at how strong this on-device model is,” said Carolina Parada, Head of Robotics at Google DeepMind. “I would think about it as a starter model or as a model for applications that just have poor connectivity.”

According to Parada, the model was initially trained on Google’s own ALOHA robot, but has been successfully adapted to other platforms such as the Apptronik Apollo humanoid and Franka FR3 bi-arm robot. It can generalize tasks with just 50 to 100 demonstrations, allowing for rapid fine-tuning on different hardware.

Alongside the rollout, Google DeepMind is also releasing a software development kit (SDK), allowing developers to test and customize the on-device model—marking the first time a VLA model has been made available in this way.

The SDK and the model itself will be initially available to a group of trusted testers as Google continues refining its safety guardrails.

 Source: The Verge