Google DeepMind presents self-training RoboCat

A new AI model called RoboCat can teach itself new tasks without human supervision, potentially leading to a breakthrough in general-purpose robotics, DeepMind says.

RoboCat can solve tasks from as few as 100 demonstrations and improve itself from self-generated data, according to DeepMind, which is Google’s AI and machine learning division.

It said that the model only took a few hours to learn how to operate different robotic arms without human supervision.

While pre-programmed to use arms with two-pronged grippers, the model trained itself to use a more complex set of arms with three.

“RoboCat has a virtuous cycle of training: the more new tasks it learns, the better it gets at learning additional new tasks,” DeepMind said in a blog post.

It noted that it was “similar to how people develop a more diverse range of skills as they deepen their learning in a given domain.”

According to DeepMind, RoboCat learns “much faster than other state-of-the-art models” and is the first that’s capable of solving and adapting to multiple tasks – and does so across different real-life robots.

“This capability will help accelerate robotics research, as it reduces the need for human-supervised training, and is an important step towards creating a general-purpose robot,” the company said.

RoboCat is based on DeepMind’s multimodal model Gato (Spanish for “cat”), hence the feline connection in its name. Gato can process language, images, and actions in simulated and physical environments.

DeepMind has published the results of its study in arXiv, an open-access research platform.