A request like 'find the box and bring it here' hides many decisions. A robot must choose a route, pick up the object and check whether each step succeeded. On October 27, Google Cloud announced public-preview^[1] access to Gemini Robotics Embodied Reasoning 2, a high-level model that plans these steps and calls separate robot-control programs.

The calls use APIs^[2], defined interfaces between software components. The model can interpret video, choose a navigation or manipulation function and check progress. Coordinating specialized components in this way is called orchestration^[3]. Google also describes cooperation between robots, but the actual behavior still depends on their hardware and control systems.

Google reports 91.3% accuracy for locating a critical moment in video, with a mean timing error of 0.96 seconds. That is a specific video-understanding benchmark, not a general robot task-success rate. The model existed before October; the new event is its cloud availability. Network delays, disconnections and emergency stops remain practical engineering responsibilities.