Gemini Robotics

What it is

Gemini Robotics is a vision-language-action model: it takes in what a camera sees together with an instruction, and outputs actions instead of words. Google DeepMind announced it on 12 March 2025 together with a companion model, Gemini Robotics-ER.^[1] Where a chat model answers a question in text, Gemini Robotics answers with physical actions that directly control a robot.^[1] It is built on Gemini 2.0.^[1]

Two models, two jobs

The announcement covered two models with different responsibilities. Gemini Robotics-ER ("ER" for embodied reasoning) is a Gemini model with advanced spatial understanding.^[1] Its focus is spatial reasoning such as pointing and 3D object detection, and it can be connected to developers' own low-level controllers.^[1] The main Gemini Robotics model is the one that produces the actions themselves.

DeepMind reported gains for both. Gemini Robotics-ER improves on Gemini 2.0's existing pointing and 3D detection abilities by a large margin,^[1] and in end-to-end tests it reached two to three times the success rate of Gemini 2.0.^[1] For Gemini Robotics, the company said it more than doubled performance on a generalisation benchmark compared with other leading vision-language-action models.^[1]

What it can do

The tasks DeepMind highlighted are dexterous ones rather than simple movement from place to place: folding origami and packing a snack into a bag.^[1] Both call for handling small objects, dealing with material that changes shape, and chaining several moves together — the kind of fine motor control that is hard to write by hand into a robot program.

Generality is the other theme. Gemini Robotics works across robot types, including the bi-arm ALOHA 2 platform, Franka-arm setups and Apptronik's Apollo humanoid.^[1] That matters because robot hardware is fragmented, and code written for one arm rarely transfers to another.

How it was trained

Gemini Robotics was trained primarily on data from the bi-arm ALOHA 2 platform.^[1] It is also described as based on the Franka arms used in many academic labs.^[1] Building on Gemini 2.0 gives the system a general model underneath the action output.^[1]

Safety alongside capability

Together with the models, DeepMind released the ASIMOV dataset for measuring the semantic safety of robotic actions,^[1] and said the dataset will help researchers rigorously measure the safety implications of robotic actions.^[1] Apptronik, whose Apollo humanoid is one of the supported bodies, was named a partner in the work.^[1]

Where it fits

The release maps out two routes into physical AI. One is end-to-end: a single model that sees, reasons and emits actions. The other keeps a reasoning model separate — Gemini Robotics-ER handles spatial understanding and hands off to a developer's own controller.^[1] Offering both, plus a shared safety benchmark, positions the work as much as an infrastructure layer for robot developers as a single model release. The claim DeepMind makes for it is measurable rather than open-ended: better generalisation than other leading vision-language-action models on its benchmark, and a large improvement in spatial abilities over Gemini 2.0.^[1]