Particle.news
Download on the App Store
6 ARTICLES
·
yesterday

Google DeepMind Debuts On-Device Gemini Robotics Model for Cloud-Free Autonomy

Built for low-connectivity and high-security settings, the model adapts to new tasks with as few as 50 demonstrations.

In a robotics lab, Google’s humanoid robot A11 demonstrates task-ready posture in front of assorted items, representing the kinds of real-world scenarios its offline AI model is built to handle autonomously
Image
Image
Apollo robot

Overview

  • Robots running the on-device model can operate entirely offline, making them suitable for environments with poor network access or strict data controls.
  • The standalone model delivers performance nearly on par with the hybrid cloud-assisted version on dexterous tasks requiring fine motor skills.
  • Google trained the model on its ALOHA robot and successfully adapted it to platforms such as Apptronik’s Apollo humanoid and the Franka FR3.
  • A software development kit is available to trusted testers, enabling developers to customize and fine-tune the on-device model.
  • Because the on-device release lacks a built-in safety layer, Google advises integrating the Gemini Live API and implementing low-level controllers to enforce force and motion constraints.