Career of the Month

Abstract

Roboticist After learning to program computers as a child in Austria, Matthias Scheutz developed an even greater interest in philosophy and the mind. He combines these passions in a career focused on using human cognition as a model to make machines smarter. He now serves as director of the Human-Robot Interaction Laboratory at Tufts University in Medford, Massachusetts. He anticipates a world in which robots routinely help people, understand spoken instructions, and make moral choices. Robotics is a huge growth area, but this is only the beginning, he says. [ILLUSTRATION OMITTED] Work overview. I focus on programming robots with intelligent algorithms and studying interactions between robots and humans. My lab does experiments in which humans and robots do tasks together so we can see how well the robots do and how well the interaction works. We also study how the people work so we can improve our algorithms. Artificial intelligence doesn't necessarily entail doing things the way a human would, but since humans are the most intelligent system we know, we look at humans as inspirations. There are two approaches to artificial intelligence: One aims to get a job done regardless of how humans would do it, while the other tries to follow the human way, which is called cognitive modeling. We use both approaches. We want to make robots intelligent so they can genuinely help people, be easier to interact with, and learn on the fly. We want machines to take spoken instructions as people do. Specific systems we're working on include a robotic wheelchair and robots that can be used in dangerous situations such as entering buildings about to collapse or working in radioactive areas. This work involves a mix of problem analysis, problem solving, and solution implementation. The engineering design process has many iterations--each time, we analyze why something failed and try to correct it. Sometimes we spend long nights or even weeks chasing errors. It's instructive, though frustrating. The research is often complex; giving robots the ability to understand language, for example, involves a lot of things working correctly. We've looked at how children learn to associate sounds with objects, and then we built a biologically plausible model of that process on a small robot. We did the same studies with the robot that developmental psychologists do with infants, for example, wiggling a new object in front of the infant and saying the object's name at the same time, and we replicated their results. We can also look inside the robot and inspect its program to see how it learns. We then expose the robot to different conditions, use that data to make predictions of how kids would respond in those conditions, and then see if we were right and whether the model is correct. …