Toyota Research Institute builds a "robot kindergarten" and uses AI to teach robots to make breakfast.

CTOnews.com, September 20 (Xinhua) Toyota Research Institute (TRI) uses generative AI to teach robots how to make breakfast, or at least the tasks needed to make breakfast, in a "robot kindergarten". Instead of spending hundreds of hours coding and debugging, they give robots a tactile sensor, plug them into an AI model, and show them how to do it, just like a human.

The researchers say tactile perception is "a key driver". By installing a "big and soft thumb" for the robot, the model can "feel" what it is doing and get more information, which makes some difficult tasks easier than vision alone.

Ben Burchfiel, the dexterous operations manager of the lab, said it was "exciting" to see them interact with the environment. First, a "teacher" will demonstrate a series of skills, and then in "a few hours", the model will be learned backstage. "We often teach a robot in the afternoon to learn at night, and then we can see a new behavior the next morning," he added. "

The researchers say they are trying to create a "large behavior model" (LBMs) for robots. Similar to LLMs, which is trained by paying attention to patterns in human writing, Toyota's LBMs will learn through observation, then "generalize" and implement new skills that they have never been taught.

Using this method, the researchers say they have trained more than 60 challenging skills, such as "pouring liquids, using tools and manipulating deformable objects," and they hope to increase that number to 1000 by the end of 2024.

CTOnews.com notes that Google is conducting similar research, similar to the approach used by Toyota researchers, whose robot RT-2 uses its own experience to infer how to do things. In theory, AI-trained robots can eventually perform some tasks without too many instructions.