Specialists from Toyota Research Institute (Tri) taught cars from the “Kindergarten for robots” how to prepare breakfast. Instead of spending a lot of time on programming and error elimination, artificial intelligence was introduced into the development process, allowing the robots to interact with the world around them. People were involved in the learning process to show the necessary movements. [1]
According to researchers, the inclusion of touch in the learning process enables AI to perform tasks that would be difficult for it to accomplish through visual analysis alone.
Bin Berchfil, the head of the “deep manipulation department” in the laboratory, commented, “It is good to observe how they interact with the environment. First, the ‘teacher’ demonstrates a set of skills, after which the model independently trains and analyzes the information received within a few hours. For us, the process looks like this: in the afternoon we train the robot, then leave it overnight, and the next morning we already see new behavioral patterns.” [1]
The process is detailed in a video on YouTube.
Scientists are also working on the creation of “Large Behavior Models” (LBMs). According to Russ, a professor at MIT, cars will be able to solve new problems even if they were not specifically taught how to do so. The robots already know how to perform more than 60 diverse and complex actions, such as pouring liquids into different vessels and handling kitchen tools and fragile objects. The goal for the near future is to expand their skills to a thousand, which would be a significant breakthrough for robotics. [1]
Similar research is being conducted by Google and Tesla. However, developing such technology is a challenging task. Unlike basic AI models that rely on mass data from the Internet, autonomous robots require a special approach. [1]
| Source: | Toyota Research Institute |
|---|