Google brings Gemini Robotics ER 2, DeepMind's embodied reasoning model, to Google Cloud in public preview so enterprises can orchestrate VLA models and multi-robot teams

Background

For robots to do real work in factories, warehouses and labs, the hard part is often judgment rather than any single motion: did this step succeed, what comes next, and how to recover when something goes wrong. Earlier this year Google DeepMind introduced Gemini Robotics 2, and its Gemini Robotics Embodied Reasoning 2 (ER 2) model handles exactly this kind of high-level reasoning; until now it was offered mainly through the Gemini API and Google AI Studio. On Oct 27 the official Google Cloud Tech account, together with a Google Cloud blog post co-written by DeepMind robotics head Carolina Parada, announced that ER 2 is now in public preview for enterprises on the Gemini platform.

What the model does

According to the blog and documentation, ER 2 is a vision-language model for robotics with the model ID gemini-robotics-er-2-preview, built on Gemini 3.5 Flash. It takes text, image, video and audio input, with a 131,072-token input limit and a 65,536-token output limit. It reasons over continuous video in space and time and classifies task progress as a binary status, discrete categories or a continuous score, so a failed step can be retried without restarting the whole workflow. Google says it pinpoints critical moments, such as when to stop pouring a liquid, with 91.3% accuracy and a 0.96-second mean absolute frame distance. It can also read values from more than 10 instrument types, including circular dials, sight glasses, digital displays, rulers and liquid thermometers.

A brain that orchestrates tools and robot teams

ER 2 is designed as a robot's high-level planner. Developers declare VLA (vision-language-action) models, navigation APIs or custom functions as tools and stream video, audio or text into the model, which plans the steps, calls the tools and self-corrects. Google says it works across stationary arms, wheeled rovers, quadrupeds and humanoids, and lets different robots share an understanding of the workspace, divide labor and hand off tasks. Operators can redirect it in plain language, and the robot can explain its approach while acting.

Enterprise tooling and early users

The point of the Google Cloud launch is the enterprise layer: IAM, security controls and built-in function calling, simulation with MuJoCo-Warp and NVIDIA Isaac Sim, curating demonstration videos in BigQuery and Cloud Storage to train custom VLA models, and Google Distributed Cloud for latency-sensitive work like SLAM and path planning on the factory floor. Per Google Cloud's documentation, the Cloud preview supports GenerateContent but not the Live API. The blog quotes early users including Boston Dynamics (video inspection with the Spot robot), Arm and DeepReach. This is a preview under Google's Pre-GA terms, and all figures are Google's own, with no independent replication yet.