When you ask a coffee-carrying robot not to spill it on your computer, you expect it to protect your nearby phone as well, without further warning. This understanding of intent is what a new research system seeks to add to the way robots learn.
According to MIT, researchers from the institute and New York University developed the Flex system, which uses artificial intelligence to derive broader rules than human instructions. In a simulation, a warning about the computer prompted a virtual robotic arm to adjust its path to avoid other electronic devices on the table.
The system works in three steps: adding the relevant elements to the task, generalizing the warning''s intent, and then eliminating information that has lost its relevance.
Thus, a warning against dropping a heavy object on a cup can be transformed into a rule that includes breakable objects.
The tests included serving coffee, carrying a candle, and holding a bowling ball. Twelve volunteers also participated in a study in which they wrote corrections to the robot''s behavior that they saw in the simulation. The team also used motion demonstrations compiled using a real arm.
The research aims to make it easier to deal with home robots and to help by reducing the need to explain each rule in detail, but the results do not mean that a home robot with these capabilities will be available, as the system''s generalizations still need wider testing in everyday life conditions