Robots are great at doing the same thing over and over again in a controlled environment. They are less useful in situations where they need to adapt. Sylvain Calinon and Leonel Rozo, two researchers at the Italian Institute of Technology’s department of advanced robotics in Genoa, are trying to change that.
In a paper to be presented at the conference of the Association for the Advancement of Artificial Intelligence in Washington state later this year, they explain how a robot can learn a task by being shown how to do it.
Their experimental robot differs from earlier robots that can learn by demonstration in two important ways. First, it can not only learn from humans, who are unpredictable and messy, but then work together with them on a task. That is more impressive than dealing with inanimate, predictable objects like conveyor belts. And its learning isn’t confined to the needs of a single person—it can adapt itself to different people, a messier proposition yet. To test it out, the researchers set their robot the task of helping humans build a table from IKEA.
Here’s how it works: humans demonstrate something several times, invariably with degrees of inaccuracy. Sensors tell the robot where it is, where the elements of the table are, and where the humans are. That means it is aware of its own position in space, much as you know where your arms are even with your eyes closed. The algorithm filters the masses of data fed to it by the sensors and extracts relevant patterns. The inspiration for the method came not from robotics but from mathematical models used in statistics, Dr Calinon told Quartz from his office in Italy.
What the researchers have planned for their robot is even more ambitious. Dr Calinon says they want to extend the machine’s abilities so it can take the lead in a task. For instance, if the person with the table leg is too far away from the robot, it should be able to move towards him on its own. It sounds revolutionary now. But perhaps one day you’ll be able to buy one at IKEA.