Google offers its robot 'brain' model to businesses in public preview
Gemini Robotics ER 2 plans multistep tasks and checks whether each step worked, the company says.
Mountain View-based Google has opened its Gemini Robotics Embodied Reasoning 2 model to businesses in public preview through Google Cloud, the company said in an Oct. 27 blog post.
The model acts as a robot's high-level "brain": it watches video, plans multistep tasks and directs lower-level software that moves the robot's arms or wheels, according to the post by Carolina Parada, head of Google DeepMind robotics, and Adrian Jones of Google Cloud. Google said it works with robots of many shapes, from fixed arms to humanoids.
Google said the model can judge whether a step succeeded and retry a failed step without restarting the job. It said the model finds the exact video frame of a key moment, such as when to stop pouring a liquid, with 91.3% accuracy, and can read more than 10 kinds of gauges and displays. Those figures are the company's own; the post did not cite independent tests.