WPI Research is the research magazine of Worcester Polytechnic Institute. It contains news and features about graduate research in the arts and sciences, business, and engineering, along with notes about new grants, books, and faculty achievements.
Issue link: http://wpiresearch.epubxp.com/i/90696
Sonia Chernova, right, and computer science graduate student Russell Toris work in the "apartment" where the robot Toris is adjusting will be controlled by online users 24 hours a day. The users' experiences will help Chernova design algorithms that will permit everyday users to train robots. is that you have to design the system from the beginning with the end user in mind," says Chernova, who is director of the Robot Autonomy and Interactive Learning (RAIL) lab. "Expecting every single user to become an expert robotics programmer is unrealistic. We want to create an interactive learning system that will work for someone with no prior robotics experience." To develop a new way of teaching robots, Chernova is pioneering a new way to do robotics research. She notes that roboticists and students are often pressed into service to test software for robots, but their prior knowledge can skew results. Instead, with a five-year, $500,000 CAREER Award from the National Science Foundation, she will bring thousands of users, from all over the world, into the testing process — via a web-based interface. Her RobotsFor.Me system allows researchers to "crowdsource" real-time trials of robotics software, to generate a large volume of data from a wide variety of people. The goal is to produce an open-source framework, for that will enable developers to rapidly and efficiently test algorithms and interfaces on a massive scale. During the study, users can log on from their home computers at any time of night or day and take a turn at remotely guiding a robot through simple household tasks, such clearing a table, or some light assembly functions that might be used in manufacturing. Users will train first on an onscreen simulation and then advance to the real thing — controlling a physical robot that is monitored by cameras. Each task is structured much like a video game, with re- wards for achieving progressive levels of difficulty. "The robot has its own apartment," Chernova jokes, "but we had to childproof its room. Our test environ- ment — which features simple, modular furniture and basic household utensils — is much more complex and less con- strained than those employed in typical user studies. As a result, we are not only developing algorithms for learning in complex places, we are also creating new techniques for keeping the robot and furniture safe." By compiling and analyzing data from a large number of diverse users, Chernova will gain a better understanding of how robots learn, and how humans typically go about instructing them. Human input tends to be "noisy" and in- consistent, she explains. Chernova is looking for better ways to structure policy algorithms, so that robots can generalize and learn from the smallest number of interactions. [7]