Gemini Robotics ER 2 is in public preview on Google Cloud

Google put Gemini Robotics ER 2 in public preview on its Gemini platform in Google Cloud. It is a Gemini model that works as the high level brain of a robot. It plans the steps of a task and uses video to see if each step worked. Companies can use it with the access and control tools of Google Cloud. The figures come from Google.

How it works

Developers declare the control software of a robot as tools. The model then sends commands to this software from natural language. It reads video, audio or text. A robot can do a step again that did not work, without a start again of the full workflow.

The numbers

Google says that the model finds the frame of an important event with 91.3 percent accuracy. The mean error is 0.96 second. It operates at 4 times the speed of many larger models, at a fraction of the cost. It reads values on more than 10 types of instrument, such as dials and thermometers.

What is not known

The text gives no test by others and no error rate in use in the field. Customers such as Boston Dynamics and DeepReach give supportive quotes, but they give no test data.

Insight: Google puts robots under the same controls as software agents

The Gemini Agent of Google starts with one point for identity, audit and spending rules for software agents. At this time, ER 2 brings enterprise access and protection controls to physical robots. This reads as a bid to make Google Cloud the control point for robots too, and the owner of that point can collect a toll. For leaders, the cloud that you select can set how robots are governed.