Google Cloud opens a preview of Gemini Robotics ER 2 for embodied reasoning

Google Cloud made Gemini Robotics ER 2 available in Preview.

Google Cloud added Gemini Robotics ER 2 to its Gemini Enterprise Agent Platform in preview on October 27. The release notes describe a vision-language model based on Gemini 3.5 Flash. The preview covers spatial reasoning, video understanding, code execution and task orchestration for robotics applications. The cloud release gives developers a model endpoint to evaluate while they build the rest of a robot application. [Google Cloud release notes](https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes#October_27_2026) · [Model documentation](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/robotics-er-2) · [DeepMind model card](https://deepmind.google/models/model-cards/gemini-robotics-er-2/)

Where reasoning fits into a robot stack

A developer can use a vision-language model to interpret a scene, answer a question about object relationships or propose a sequence of operations. In a hypothetical inspection workflow, the reasoning layer might identify the part of a machine that a camera should inspect next. Motion control, collision checks and the decision to stop the equipment remain separate system responsibilities. The cloud availability of the reasoning model does not certify the full machine or its operating environment.

The distinction also shapes what to evaluate. A static image question can test whether the model identified the right object. A video test can assess whether it noticed a change as the scene evolved. An orchestration test must check whether it chose the right permitted tool and handled a failure. Each result says something different about readiness. [DeepMind model documentation](https://deepmind.google/models/gemini-robotics/embodied-reasoning/)

What preview status means for testing

An evaluation can begin with logged camera feeds and simulated actions, where teams can replay failures and compare predictions with known outcomes. They can then measure how often incomplete or ambiguous observations cause the model to choose the wrong object, skip a required step or request more information. Measuring response latency matters too, especially when a scene changes while a cloud request is in flight.

The availability milestone is specific to the documented preview endpoint. The production decision remains with the team integrating perception, reasoning and motion. Its evidence needs to cover that entire application, including how the robot behaves when the model is unavailable.