AI in robotics: Toyota Research Institute's "Robot Kindergarten" program

In some previous news we learned that Google was working on theAIResearch on applications in robotics. Google's related robot, RT-2, can use the experience it gains to infer how to do things. In other words, AI-trained robots may eventually be able to perform some tasks without much instruction. The development of such technology will undoubtedly bring more convenience and possibilities to our lives.

Toyota or Toyoda (name)Research Institute (TRI) recently revealed that they are utilizing generativeAI technologyIn a pioneering project called Robot Kindergarten, robots are being taught tasks such as making breakfast. At the heart of this innovative approach is the idea that by giving a robot a tactile perceptron and plugging it into an AI model, researchers can show the robot how to perform tasks directly, without having to spend hundreds of hours coding and debugging.

Haptic perception is seen as a key enabler in this process. By giving the robot a "big, soft thumb", the model can "feel" what it is doing and thus gain more information. This makes tasks that would normally be considered difficult, such as pouring liquids, using tools and manipulating deformable objects, much easier to accomplish than vision alone.

Ben Burchfiel, dexterity manager at Toyota Labs, is excited. He describes a scenario in which a "teacher" shows the robot a range of skills, and then within a few hours the model is learning itself in the background. He adds, "We often teach a robot in the afternoon, let it learn at night, and then see a new behavior the next morning."

To train robots, Toyota researchers are trying to create "Large Behavior Models" (LBMs). Such models are trained in a similar way to natural language processing models (LLMs), in that they learn by observing human behavior. However, Toyota's LBMs will go a step further by learning through observation and then "generalizing" to perform new skills they have never been taught.

So far, the researchers have successfully trained more than 60 challenging skills, such as pouring liquids, using tools and manipulating deformable objects. Their goal is to increase the number of skills they can train to 1,000 by the end of 2024.