Researchers deepmind taught robots to learn new skills at an astonishing rate. Representatives of the company reported this in an article published on arXiv.

They have developed an AI called RoboCat that uses the same technology as large language models (LLM) is the architecture of the transformer. This technology allows AI to process huge amounts of data and use its knowledge to perform new tasks.

RoboCat has been trained on tens of thousands of demonstrations of four different robotic arms performing hundreds of different actions, such as sorting colored bricks or picking the right fruit from a basket. These demonstrations have been produced both by humans controlling the robots from a distance, and by special AIs controlling virtual robots.

Through this training, RoboCat has become a generalist capable of handling a wide range of robotics tasks using a variety of hardware configurations. In addition, he can now learn new tasks by watching 100 to 1000 demos from a human robot. This is much less than what is typically required for training, indicating that the model is building on more general robot control skills rather than starting from scratch.

The most interesting thing is that RoboCat can improve itself. The researchers created several modified task-specific models and then used these models to generate approximately 10,000 new task demonstrations. These demos were then added to the existing dataset and used to train a new performance-enhancing version of RoboCat.

When the first version of RoboCat was shown 500 demonstrations of a new task for the robot, it could successfully complete it 36 ​​percent of the time. But after several rounds of self-improvement and self-learning, that figure doubled very quickly to 74 percent.

In solving some problems, the model still demonstrates success below 50 percent, and for one specific problem it is even 13 percent. However, the ability of RoboCat to master many different skills and quickly pick up everything suggests that adaptive intelligence for robots is just around the corner.