Particle.news

Download on the App Store

Google DeepMind Launches Gemini Robotics On-Device for Cloud-Free Robot AI

The model brings vision-language-action inference to robots without internet access by running locally with low latency.

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
Jakarta - July 26,2024: Google DeepMind, formerly DeepMind Technologies logo signboard, is a British artificial intelligence research laboratory which serves as a subsidiary of Google
Image

Overview

  • Gemini Robotics On-Device operates entirely offline on a robot’s onboard hardware, achieving performance near DeepMind’s cloud-based model.
  • Developers can fine-tune the model with natural language prompts after providing just 50 demonstrations for new tasks.
  • The release includes a Gemini Robotics SDK that supports simulation evaluation; it enables deployment across diverse robotics platforms.
  • DeepMind demonstrated cross-platform adaptability by running the model on ALOHA robots; it was later adapted to a bi-arm Franka FR3 and Apptronik’s Apollo humanoid.
  • The on-device variant omits built-in semantic safety reasoning, prompting developers to replicate DeepMind’s safety stack with custom safety controllers.