Many big factories have robots working in them, as well as humans, but the two don't often work together. At MIT, researchers have come up with an algorithm that may help solve that problem and make it easier and safer for humans and robots to work side-by-side.
Julie Shah, who leads the Interactive Robotics Group in MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) and her colleagues devised a way that gives robots the tools to learn the preferences of a human coworker.
Why bother? Many jobs need both robots and humans. A robot is really good at repetitive, simple tasks. But fine work like refinishing or even placing objects correctly can stump it, because machines aren't good at certain kinds of thinking, and often are quite clumsy compared to humans. In addition, people might do the same job in different ways, and robots are notoriously bad at understanding that. So the usual procedure is for a robot to do one thing and a human to do another, but separating the two tasks isn't as efficient as having the two done at the same time.
The experimental set-up Shah established involved making airplane parts. One step requires a mechanic to add sealant to predrilled holes, hammer bolts into the holes, and wipes away the excess sealant. But different mechanics like to do this in their own ways; one might apply sealant to every hole before hammering the bolts, another might finish one hole-and-bolt combination before doing the next one.
Using a robot called FRIDA, designed by the Swiss company ABB, the research team dressed a worker up in the same kinds of motion-capture suits used in movies. "Kinect is great," Shah told Discovery News, "But we needed something more precise." The suit gathered data that was shown to the robot in a virtual environment to communicate how each human did the work. The idea is to help the robot decide: do I apply sealant, doing each hole in sequence, or place the bolt in immediately after applying it?
The robots quickly adapted, either applying sealant or placing bolts depending on how each human mechanic liked to work.
The system isn't quite ready for factories yet, Shah said. Obviously using motion capture suits to pre-train the robots isn't practical for a working factory. So there needs to be a more seamless way of teaching the robot. RFID chips might provide part of the answer; one could imagine tagging a person with them so a computer could track their movement.
It's also important to design robots that can be inherently safe for humans to work around. Since robots can't "see" a person, it's all too easy for accidents to happen (this is one reason robots work separately from people in the first place). Robots could be made lighter and with sensors that show sudden increases in torque (which would signal that, say, a rotating arm has hit something).
Shah added that the work also offered insights into how to do a better job of teaching machines. For example, most current machine learning algorithms are simple reinforcement: if something is done right the robot gets a "yes" signal. It turned out that having the robot do different jobs in the same process was also helpful. Cross training in this way is also used in human manufacturing teams, because someone that knows how to do partner's job is able to more efficiently work with that person.
Some worry that robots will completely replace humans, but Shah said even with systems like this that isn't likely to happen. There are too many areas that human judgment is needed, and robots are still really bad at that if the task isn't very specifically outlined. But the day will soon come, she said, when robots and humans can work alongside each other with the robot learning seamlessly how each person works. And that cold change the way we program them. "For that kind of learning the formalism doesn't exist in programming," she said.
Top Photo: ABB industrial robot in Villeneuve-d'Ascq, France