Google DeepMind Launches Gemini Robotics On-Device Model for Offline Robot Control

Google DeepMind has introduced Gemini Robotics On-Device, a new language model designed to run locally on robots, enabling them to perform complex tasks without needing an internet connection.

Gemini Robotics On-Device is a lightweight robotics foundation model for bi-arm robots, enabling low-latency, on-device task execution.

It supports natural language commands, adapts through fine-tuning, and excels at dexterous tasks like folding clothes or unzipping bags, demonstrating strong generalisation across visual, semantic, and behavioral domains in real-world scenarios.

Google claims its performance is close to its cloud-based counterpart and superior to other on-device alternatives—though it hasn’t specified which models those are.

"Since the model operates independent of a data network, it’s helpful for latency sensitive applications, and ensures robustness in environments with intermittent or zero connectivity," Google said in a blog post.

In live demonstrations, robots powered by the model successfully performed tasks like unzipping bags and folding clothes.

Initially trained for ALOHA robots, the model was later adapted for the Franka FR3 dual-arm robot and Apptronik’s Apollo humanoid, where it managed unfamiliar tasks such as industrial assembly.

Google also announced a Gemini Robotics SDK to support training. Developers can teach new tasks using 50–100 demonstrations in the MuJoCo physics simulator.

During Google Q1 2025 earnings calls, Alphabet CEO Sundar Pichai revealed that Google is working on a family of Gemini models for robotics.

"Lastly, we're developing AI models in new areas where there’s enormous opportunity, for example, our new Gemini Robotics models," he said.