The preview brings a robotics reasoning model into Google Cloud
Google Cloud announced public-preview access to Gemini Robotics ER 2 on October 27. The vision-language model accepts text, images, audio and video, then returns text or images for tasks such as spatial reasoning, instrument reading, progress classification and multi-step tool orchestration. Developers can connect lower-level robot controls or navigation functions as tools.
This is an availability change for a model Google introduced earlier in 2026, not a claim that customers receive a complete robot-control system. The model can propose plans and call declared interfaces, while the customer remains responsible for the hardware, control layer, permissions and operational safeguards. CGH has not tested the model's accuracy, latency or behavior on physical equipment.
Access is global, with material preview limits
Google lists the model ID as gemini-robotics-er-2-preview and makes it available through a global endpoint under Standard PayGo consumption. The default project quota is 250 requests per minute, with a documented range of 200 to 300 based on quota tier. The reviewed model page does not state a robotics-specific dollar rate, so customers should confirm the applicable input, output and media charges before a production budget is set.
The model supports function calling, structured output, code execution and several grounding tools. It does not support tuning, batch inference, provisioned throughput, fixed quota, the Gemini Live API or the Chat Completions API. Google also labels it a pre-GA service that may have limited support; an example deployment requires a billed Cloud project and the Agent Platform API.
Physical trials need stricter evidence than a software demo
A useful evaluation should begin in simulation or with a constrained machine, a low-energy workspace and a tested stop mechanism. Teams should log every model request, tool call and operator intervention, then measure failure recovery against defined tasks rather than vendor demonstrations. Before any live deployment, they should confirm data-location requirements, preview support, rate limits, cost controls, human approval points and responsibility for unsafe commands. Google's own guidance says these models can make mistakes and that people remain responsible for robotic safety.