Gemini Robotics ER 2 reaches public preview on Google Cloud

An existing model gets a wider Cloud entry point

Google opened Gemini Robotics ER 2 to public preview on Google Cloud on October 27, giving developers and enterprise teams a broader route to its model for interpreting scenes and planning robot tasks. The release makes it available through Model Garden with Cloud identity controls and project-based access.

The change is in access, rather than a new model generation. Google introduced ER 2 on July 30, when its announcement described public availability through the Gemini API and Google AI Studio, alongside a private preview on the enterprise platform. Today's release opens that previously limited Cloud route.

For organizations already managing robotics infrastructure in Google Cloud, the practical attraction is bringing this reasoning component into their existing cloud environment. The preview still leaves the robot's control software and physical safeguards to the deployment team.

Source 1Source 2Source 3

From a visual instruction to calls into robot software

ER 2 works above the machinery that moves a robot. It interprets the scene, decides which steps to attempt and calls tools supplied by the developer. Google's July description explicitly separates this planning role from motor execution handled by lower-level action models.

The Cloud task-orchestration guide makes that boundary concrete. Its example asks for a blue block to be placed in an orange bowl, exposing functions for arm movement and opening or closing a gripper. An application loop executes requested calls, returns their results to the model and continues until calls stop or a step limit is reached.

This is a mock robot API: the supplied functions print actions. The example demonstrates how software connects a model's decisions to controls; it does not demonstrate a successful physical pickup. Turning it into a working installation requires implementing those functions for actual hardware.

Source 2Source 7

Video analysis and live streaming are different capabilities

The video guide describes two useful checks: finding the time a task finishes and estimating how far it has progressed. Its completion example asks for a timestamp or a null value if the task remains unfinished. Another example assigns the final video frame to one of five progress brackets. Both send video files through a request, illustrating how an application can ask about observed work.

That does not establish a live streaming connection on Cloud. The Cloud model specification supports GenerateContent but lists Gemini Live API as unsupported. By contrast, the Developer API documents a separate gemini-robotics-er-2-streaming-preview endpoint for continuous audio and video. A developer moving a streaming application to Cloud therefore cannot assume the same interface is available.

Source 8Source 4Source 5

Cloud credentials, a global endpoint and preview conditions

The Cloud setup guide uses Google Cloud credentials rather than a Gemini API key. Its SDK examples select a project and the global location. The model page also requires billing and the Agent Platform API to be enabled for its example-app deployment.

The preview's listed serving location is global, with no regional endpoint offered in the documentation. Standard pay-as-you-go consumption is supported; batch inference, provisioned throughput and model tuning are not. Those restrictions matter when choosing how to run repeated video analysis or reserve capacity.

Google's documentation permits production or commercial use subject to the applicable agreement, while warning that pre-general-availability features are supplied as is and may have limited support. Public preview should not be read as general availability.

Source 3Source 4

The introductory prices depend on credits

Google's pricing table lists ER 2 input at $1 per million tokens and text output, including reasoning, at $5 per million through December 31, 2026. Cached input is listed at $0.10 per million. The table's footnote says promotional pricing is provided through 50% credits back on net spend on selected models within a given period.

That mechanism matters for budgeting: the displayed promotional rates should not be described as an unconditional upfront discount. Starting January 1, 2027, the table lists $2 for input, $10 for output and $0.20 for cached input. These are model-token prices, not the total cost of operating a robot or its surrounding infrastructure.

Source 6

Deployment responsibility stays with the operator

The overview warns that the model can produce incorrect answers and that higher-resolution inputs or increased reasoning can add latency. It places responsibility for a safe operating environment on the robot operator.

Its privacy notice also requires sufficient notice and consent before identifiable people interact with, or are present around, a deployment that collects their personal data. It calls for reasonable efforts to minimize that collection, including face blurring where practicable. Those requirements make camera placement and who can enter the workspace part of the implementation, alongside the API integration.

Source 3