Universal robots are no longer far away. Google demonstrates the world's first multi-tasking AI agent

Recently, Google's AI team DeepMind launched an AI agent for robots that can improve and enhance itself, named RoboCat.

RoboCat is essentially an AI-powered software program that can serve as the "brain" of a robot. The difference between robots powered by it and traditional robots is that RoboCat robots are more "versatile" and can achieve self-improvement and self-enhancement.

DeepMind said that RoboCat is the world's first robotic AI agent that can solve and adapt to a variety of tasks, and it can complete these tasks on various real robot products.

According to DeepMind, with only about 100 demonstrations, RoboCat can learn to control the robotic arm to complete a variety of tasks, and then iterate and improve through self-generated data.

You know, one of the important reasons why the construction of general robots has been slow is that it takes time to collect real-world training data. RoboCat's rapid learning ability reduces the need for human-supervised training, which can be said to be an important step towards creating a general robot.

From the released video, we can see that RoboCat can already control the robotic arm through autonomous learning to complete tasks such as "ringing", "building blocks", and "catching fruits". These tasks may seem simple, but they test the accuracy of the robotic arm's operation, its comprehension, and its ability to solve shape matching problems.

Most importantly, RoboCat has never seen either the robotic arm it controls or the task it has to complete before. Now, RoboCat's success rate in completing a new task has increased from 36% in the early days to 74%.

One of the key technologies used by RoboCat is a multimodal model called Gato. Gato means "cat" in Spanish, which is one of the origins of the name "RoboCat".

The Gato model can process language, images, and actions in both simulated and physical environments. The researchers combined Gato's architecture with a large training dataset containing 100-1000 demonstrations of various robotic arms completing tasks.

Based on the original dataset and the data generated by the new training, RoboCat's dataset will contain millions of training trajectory data. The more new tasks it learns, the better it will be able to learn and solve additional new tasks.

Researchers have previously explored robots learning multiple tasks at scale and combining their understanding of language models with real-world robotic capabilities, but RoboCat is a breakthrough in that it is the first robotic AI agent that can solve and adapt to multiple tasks.

DeepMind's paper shows that the significant improvement in the success rate of task execution is due to RoboCat's increasing experience, just as people develop more diverse skills when they deepen their learning in a specific field. In addition, RoboCat's success rate in completing real-world training tasks is much higher than that of traditional vision-based model solutions, which is also an important value of DeepMind's research.

RoboCat's "universal learning ability" is of great significance for accelerating research in the field of robotics. DeepMind believes that RoboCat's ability to independently learn skills, quickly improve itself, and quickly adapt to different hardware devices will play an important role in promoting the development of a new generation of general-purpose robotic AI agents.