The History of AI

Gemini Robotics

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Images, video, and the physical world

Lineage
Acting in the world
Kind
Implementation

Who

Google DeepMind

What happened

Google DeepMind presents Gemini Robotics, a vision-language-action model built on Gemini 2.0 that controls robots directly, and Gemini Robotics-ER, with spatial reasoning, for roboticists to use in their own programs.

Why it mattered

A frontier model controls physical robots: actions become one more kind of output, alongside text and images.

The full story

Frontier language models understood text, images, and video, but they produced only text or images. Gemini Robotics added physical actions as a new kind of output, to control robots directly.

Google DeepMind described it in terms of three qualities. Generality: it handles varied tasks without prior training, including some it has never seen. Interactivity: it continuously monitors its surroundings and adjusts its actions if objects or instructions change. Dexterity: it completes multistep tasks that require precise manipulation, such as folding an origami figure or packing a snack into a zip-top bag.

It was trained mainly on data from ALOHA 2, a two-armed robotic platform, but it also controls Franka arms and can be specialized for Apptronik's Apollo humanoid robot. The technical report adds that, with fine-tuning, it learns new short tasks from just a hundred demonstrations. Gemini Robotics-ER was offered to companies such as Boston Dynamics and Agility Robotics, and Google released the ASIMOV dataset for measuring the safety of robot actions.

Its creators presented it as a path toward general-purpose robots, built on the same foundation as Google's language models.

Videos

Sources

Original source

How to cite

“Gemini Robotics.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/gemini-robotics

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline