Researcher Founds a Robot Soccer Dynasty
Since receiving her doctorate in 1992, Manuela Veloso's research interests in artificial intelligence have focused on duplicating the success with which humans plan, learn and execute tasks. Founding a robot soccer dynasty was purely coincidental.
March 24, 2004
Editor's note: Nearly 300 teams from 37 countries are expected to compete in RoboCup 2007 in Atlanta, Ga. The international research robotics competition runs from July 1-10, 2007.
Since receiving her doctorate in 1992, Manuela Veloso's research interests in artificial intelligence have focused on duplicating the success with which humans plan, learn and execute tasks. Founding a robot soccer dynasty was purely coincidental, but a logical result of putting her research into practice with her students.
Veloso and her students in the multi-robot CORAL Lab (Cooperate, Observe, Reason, Act, and Learn) at Carnegie Mellon University in Pittsburgh have pursued research to program a robot -- and teams of robots -- to make plans, carry them out, and learn from their successes and failures. Not to mention scoring enough goals to win a few robot soccer world championships in the process. Veloso's work has been supported by a 1995 CAREER award and a subsequent 1999 award from NSF.
A vice president of the international RoboCup Federation, Veloso co-chaired the first RoboCup American Open in 2003. In addition, Veloso's students, former students and post-docs are leading or participating on robot soccer teams from Carnegie Mellon, Georgia Tech, the University of Texas at Austin and, at the international level, from the University of New South Wales in Australia.
"Because of my research interests, I became attracted to multi-robot problems, and test platforms are hard to come by for such problems," said Veloso, a professor of computer science and director of the CORAL Lab. "Robot soccer came later, and it made team problems all the more challenging from a research perspective. What drives me is the multi-robot aspect and the adversarial game domain. It's kind of like chess, but with robots."
RoboCup is an international project to foster advances in artificial intelligence and intelligent robotics research. The ultimate goal of RoboCup is to develop, by 2050, a team of fully autonomous humanoid robots that can beat the human world champion soccer team. Veloso and Carnegie Mellon have been participating since "pre-RoboCup" events in 1996 and the first official RoboCup games in 1997. Veloso was general chair of RoboCup 2001 in Seattle.
The American Open was designed to help teams from the Western Hemisphere prepare for RoboCup 2003 in Padua, Italy. "Other regions, particularly Europe, have had opens for several years," Veloso said. "The American Open is a chance for the teams from the Americas to bond more. My vision was to share more amongst ourselves so the Americas could perform better in the international competition."
Teams at the American Open competed in three leagues. The simulation league pit teams of 11 simulated players against one another on a virtual playing field. In the small-size robot league, teams of five six-inch robots competed on a surface the size of a ping-pong table. Each school designs their own robots within the size restrictions. The four-legged robot league featured teams of four Sony AIBO robot dogs, donated by Sony, with each team relying on superior programming to beat their competitors.
Led by Veloso, Carnegie Mellon was the RoboCup 2002 four-legged robot league champion and has been champion in each league several times—simulation champs in 1998 and 1999, small-robot champs in 1997 and 1998, and four-legged robot champs in 1998 and 2002.
"It keeps getting harder and harder to win every year," Veloso said. "My students finish their Ph.D.'s, and they end up going to other schools where they become some of our stronger competitors. And since this is a research project, we make our code freely available, so anyone can start where we were last June. In many ways, we may have to play against ourselves."
Carnegie Mellon's teams now compete against Georgia Tech and the University of Texas at Austin in the four-legged robot league, and against Texas in the simulation league. Georgia Tech's team is led by Veloso's former post-doctoral researcher, Tucker Balch, and the Texas teams are led by Veloso's former graduate student Peter Stone. At the international level, Veloso's former graduate student, Will Uther, participates on the team from the University of New South Wales in Australia, which finished second to Carnegie Mellon in the four-legged robot league at RoboCup 2002.
"I can't imagine having had a better advisor than Manuela," said Stone, now an assistant professor of computer science, whose own 2003 NSF CAREER award will be used, in part, to support his RoboCup-related research. "She provided all the resources needed to get us to, and keep us at, the forefront of the emerging international RoboCup community. She taught me how to balance research with this sort of activity, and she instructed me in all aspects of being a full-fledged participant in the scientific community."
"I'm very happy when I see a particular research solution succeed" no matter which team makes it happen, Veloso said. In 2003, for example, Carnegie Mellon evaluated different kicking methods as well as various teamwork algorithms to include a better response to the opposing team. "I refuse to make a prediction on who will win," she joked. "I do predict, however, that we'll see some key advances in the science of autonomous robots."
UPDATE: Carnegie Mellon’s teams finished in first place in the small-size robot and four-legged robot teams in the American Open 2003, but ran into stiff competition at the international RoboCup 2003. Two teams led by Veloso’s former students, however, finished well: The University of Texas at Austin took first in the online coach competition, and the University of New South Wales won the four-legged robot league. Defending champions Cornell won in the small-size robot league.
-- David Hart
#9502548 CAREER: Planning, Execution, and Learning in Autonomous Agents
#9900298 NSF-CNPq Collaborative Research: Multiagent Collaborative and Adversarial Planning, Execution, Perception, and Learning
American Open 2003: http://www.americanopen03.org
CMU CORAL research group: http://www.cs.cmu.edu/~coral/
CMU MultiRobot Lab: http://www.cs.cmu.edu/~multirobotlab/