Google Cloud has announced a public preview of Gemini Robotics Embodied Reasoning 2 for its enterprise platform. For robotics teams, choosing the right interface matters when designing an integration.

The big change

  • When an application receives a proposed action, robot-specific code must carry it out and report back. That boundary determines where teams can check an action, detect a failure and decide whether it is safe to continue.
  • The Cloud and Developer API routes have different interfaces.
  • The meaningful measure is the whole system: a correct reading of a video is only useful if the resulting action happens in time and the application can handle a failed or unexpected result. The preview opens an integration path, but it does not establish reliability for a physical process.

Two model IDs, two interfaces

The Cloud specification lists gemini-robotics-er-2-preview with an October 27, 2026 release date. It accepts text, images, audio and video, returns text, and supports GenerateContent and function calling; Live and Interactions are unsupported. The Cloud guide uses the Google Gen AI SDK with Cloud project credentials; its example application also requires billing and the platform API.

A separate Gemini Developer API streaming guide explicitly requires gemini-robotics-er-2-streaming-preview, rather than the standard ID. Its SDK example uses a Gemini API key. Its Live connection uses a persistent WebSocket, accepts audio, images and text, and returns text. The sample sends observations and receives blocking tool calls. The application executes each requested function and returns its result to the persistent session; its robot executor is a stub that the developer must replace.

An action proposed by the model is not evidence that the robot completed it: the result returned to the model needs to reflect execution, and failures need a response. The sample illustrates an orchestration loop, not a complete robot controller or evidence that the Cloud endpoint supports that streaming model.

Google's DeepMind model page lists Live support at the family level without distinguishing those IDs. It also calls enterprise access a private preview, while the Cloud announcement calls it public preview. The pages do not explain that status discrepancy.

What the integration still needs to establish

Google introduced ER 2 in July, describing public Gemini API and AI Studio access alongside a private enterprise preview. Its demonstration used a Boston Dynamics Spot with navigation and manipulator interfaces. That was Google's demonstration, not a measurement of the new Cloud route in a customer's environment.

The Cloud announcement describes video-based task assessment and reports 91.3% accuracy for finding a significant moment, with a 0.96-second mean absolute frame distance. It does not explain the evaluation sample or method alongside those figures. They do not tell an integrator how often a whole physical task will succeed, or what happens when a requested action cannot be completed.

Access and cost also depend on the route. Google's Cloud pricing and Developer API pricing are separate schedules, so teams should check the applicable model, tier and current terms before estimating a trial. BIG CHANGE has reviewed the documentation but has not run ER 2, operated a robot with it or reproduced Google's figures.