Toyota Research Institute (TRI) used generative artificial intelligence in "Robotic Kindergarten" to teach robots to make breakfast -- or at least the individual tasks required to make breakfast -- without requiring hundreds of hours of programming and bug fixing. Instead, researchers achieved this in a short time by giving robots a sense of touch, plugging them into an artificial intelligence model, and then teaching them what to do just like they would a human.

The researchers say touch is "a key enabler." By having the robot extend the pillowy thumb you see in the video below (my words, not theirs), the model can "feel" what it's doing, allowing it to gain more information. This makes difficult tasks easier to accomplish than by sight alone.

"It's exciting to see them interacting with their environment," said Ben Burchfiel, manager of the lab's dexterous operations department. First, a "teacher" demonstrates a set of skills, and then "over a few hours," the model learns in the background. "We often teach a robot in the afternoon, let it learn overnight, and then see its new behavior the next morning," he added.

The researchers say they are trying to create "Large Behavior Models" (LargeBehaviorModels), or LBMs, for robots. "Similar to how LLMs are trained by recording human writing patterns, Toyota's LBMs will learn through observation and then "generalize, performing a new skill that they have never been taught," said Russ Tedrake, professor of robotics at MIT and vice president of robotics research at TRI.

Using this process, the researchers say they have trained more than 60 challenging skills, such as "pouring liquids, using tools and manipulating deformable objects." They hope to increase this number to 1,000 by the end of 2024.

Google and Tesla have been conducting similar research with their RoboticTransformerRT-2. Similar to the approach of the Toyota researchers, their robot uses its own experience to infer how to do things. In theory, AI-trained robots could eventually perform tasks with almost no instructions, other than giving a human a general instruction (such as "clean up a spill").

But as The New York Times noted when reporting on the search giant's research, Google's bots at least still have a long way to go. This kind of work is often "slow and labor-intensive," and providing enough training data is much harder than feeding an AI model large amounts of data downloaded from the Internet.