In recent years, artificial intelligence has reshaped the digital era. Google’s AI unit DeepMind has crossed another milestone in AI technology by developing a self-learning AI tool, Robotcat AI, which can learn new tasks without supervision.
A first of its kind, as DeepMind claims, RoboCat AI can use various real-world robots to teach itself and solve multiple problems. The data from real robots allow the AI tool to improve its functioning and share it with other robotic devices.
According to DeepMind, RobotCat is a significant step towards designing robots that can perform general, daily tasks without human intervention. In a blog post about the company’s latest AI development, DeepMind researchers wrote, “RoboCat learns much faster than other state-of-the-art models.”
Per the blog post, RoboCat AI collects data from large and diverse datasets, allowing it to learn new tasks in about a hundred demonstrations. “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 post continued.
DeepMind’s AI model Gato served as a blueprint and inspiration for Robotcat, as it, too, learns by studying events, texts, and pictures. After creating RobotCat, the researchers let it train by itself and noticed as the AI tool performed a task approximately 10,000 times without human supervision, indicating rapid improvement.
The AI-driven robot learned to perform over 250 tasks across four types of robots. It also demonstrated adaptability capabilities by adapting its self-supervision training to evolve from a two-fingered to a three-fingered robot arm. As the AI model gradually developed, it could teach itself previously hidden tasks.
The advancements in RoboCat’s functioning were due to its growing experiences; the more it learned, the more it improved, similar to humans developing new and various skills as they acquire more knowledge of a particular discipline. “RoboCat has a virtuous cycle of training: the more new tasks it learns, the better it gets at learning additional new tasks,” DeepMind’s blog post read.
As the demand for domestic robots increases, DeepMind researchers continue to adapt to the trend of self-learning robots to create AI robots that can carry out general and everyday tasks.
Carnegie Mellon University engineers also have a robot in the works that can develop new skills by watching videos of humans performing them. The team built a robot capable of opening shelves and lifting knives to cut fruits, acquiring these skills within 25 minutes.