Engineers from the Department of Electrical Engineering and Computer Science (EECS) at the Massachusetts Institute of Technology decided to endow robots with “common sense” using a large artificial intelligence language model. Researchers are confident that this will make robots truly useful in housework.
Modern robots are quite effectively programmed to perform specific tasks within the limits of their physical capabilities, and they will do this safely until external conditions change. Robots cannot handle the unexpected, which is why Boston Dynamics had to work so hard to develop machines that can maintain balance. To overcome this problem, MIT engineers linked the physical movements of robots to AI models that are used to generate content.
They took an approach that allowed the robot to break tasks down into subtasks. By taking things step by step, he adapts to unexpected events without starting over. This means that engineers won't even have to program the robots with fixes to account for all contingencies. Large language models access their own data libraries to generate words, images, computer code, or any other supported content. In this project, words were replaced with subtasks.
The technology was tested on a robotic arm, which was given the task of transferring balls from one bowl to another. A large language model helped break this task down into several subtasks, mapping them to the robot's movements. The researchers allowed it to be completed, and when the robot learned the basics of performing the task, the engineers began to “disturb” it by moving the manipulator. With a traditional algorithm, the machine would have to return to a known starting point to continue the task, but the AI model helped it determine its own position in any condition and continue from where it left off.
The authors of the project are confident that this will help in creating home assistants capable of adapting to the environment and coping with external difficulties – a person will not have to program the robot for all occasions, because he will be able to learn independently through an AI model.
If you notice an error, select it with the mouse and press CTRL+ENTER.