Fair game for intuition

Designing a robotic volleyball team may be fun but does it contribute anything to the development of AI and real-world applications? Malcolm Peltu reports

Next week, teams from United Kingdom and Irish universities will fight out the Lego robot volleyball competition at the British Association's annual festival of science in Sheffield. Last month, international robot soccer tournaments were held in Sweden and Brazil. Other robot sports events are mushrooming.

Artificial intelligence (AI) and robotics researchers are adept at spinning out the public relations potential of their work, typified by the global hype when the Deep Blue computer program beat world chess champion Garry Kasparov in 1997.

A new challenge by the artificial intelligentsia could grab even more media attention: by mid-21st century to create a robot football team to beat the reigning human world champions. That is the target of RoboCup, the Japanese-inspired Robot World Cup Initiative which organised the biggest of last month's robot soccer "world championships", involving about 90 teams from 24 countries at the International Joint Conference on AI in Stockholm.

Such events may be effective ways of promoting these disciplines, but are they of any practical research value? IBM clearly thought so when it spent millions of dollars and five years building Deep Blue.

AI pioneer Donald Michie, professor emeritus of machine intelligence at Edinburgh University, says this investment brought advances in combining super-powerful search programs with large numbers of custom-built chips to solve problems with seemingly intractable levels of combinatorial complexity. These occur in many scientific areas, like bio-molecular medical research.

Professor Michie is confident team-sport robotics could make a greater impact by stimulating developments in "social intelligence" between autonomous, co-operating machine "agents".

"This is one of the most neglected areas in AI and cognitive science," Michie explains. "Ultimately, the spur of competition will result in the development of design principles and technologies for creating insightful team players who can co-operate and second-guess the intentions, objectives and needs of other intelligent machine agents."

Just think what is involved in a classic Manchester United goal, when David Beckham picks up the ball on the halfway line; dribbles past a few players; sprints down the wing; and crosses to a diving Dwight Yorke, who scores. Much more than excellent ball control and speed is needed. Players also have to make a stream of instantaneous, mainly intuitive, decisions. These knit together myriad pieces of personal and social knowledge, including team strategy, an individual's own capabilities and anticipation of the behaviour of team mates and opponents.

"Effective responses to real-time pressures like these require the graceful co-ordination of sensors, planning, learning, understanding and other factors which provide unprecedented challenges for developing intelligent and integrated software/hardware engineering," Michie points out.

Support for Michie's expectations comes from Manuela Veloso's computer science group at Carnegie Mellon University, which has competed in all three RoboCups since 1997. A member of this group, Peter Stone, argues: "The beginnings of most of my best research contributions have come in the few months leading up to the competitions. They were always solidified afterwards, but the competitions have continually motivated innovations.

"The RoboCup simulator league also played an important role in my doctoral thesis," he adds. "It provided tangible evidence that my concept of hierarchical 'layered' machine learning for multi-agent systems could outperform monolithic learning approaches used by other teams."

Stone believes a crucial reason why Carnegie Mellon won this year's RoboCup simulator league was its introduction of explicit reasoning to anticipate what team mates and opponents are likely to do, which is a vital element in modelling social intelligence.

The Carnegie Mellon team has also won the RoboCup league for small real robots and this year entered the first Sony Legged Robot League (mobile robots still generally move on wheels). "Without the stimulus of RoboCup, we would probably have continued working only in simulation," observes Professor Veloso.

The physical aspect of robotics sport is highlighted by Kevin Warwick, professor of cybernetics at Reading University: "People need to 'touch robots' in building them to fully appreciate the problem of coping with so many purely physical technical unknowns, let alone trying to integrate them with intelligent software."

Essex University's Huosheng Hu led the sole UK contender at RoboCup'99, coming a creditable third in the simulation league. "One of the important contributions of this kind of event is to provide a common research platform on which various theories, computing algorithms and system architectures can be evaluated," he says.

Such a platform is particularly important in robotics, according to Ulrich Nehmzow of Manchester University, one of the robot volleyball contestants. "Traditionally, robotics researchers focused on individual experiments in unique environments that are difficult to describe. We must move closer to rigorous scientific methods for replicating experimental results and evaluating different approaches."

Nehmzow looks for practical benefits even from the relatively simple game of volleyball: "The way a robot learns to identify the position of a ball in its half of the court could help our research into enabling mobile robots to navigate around a building, where they could encounter people and objects unexpectedly."

Working with the relatively limited infrared and audio sensors available in the robot volleyball contest could also be a productive challenge. "It could help us to learn about the cognitive processes people use to transform incomplete sensory inputs into meaningful information," Nehmzow says.

Edinburgh University's team is interested primarily in biological robotics, particularly their social interactions and the relationships between control and physical robot structure that generate biological behaviours. "Games such as volleyball emphasise situations that address the principal difficulty in robotics research, which is to design and construct reliable behaviour in dynamic and uncertain environments," says Edinburgh's John Hallam.

Gerard Lacey at Trinity College Dublin says these events helped to develop distributed-control system models and small real-time controls for mobile robots. He hopes sensory motor controllers created for robot volleyball will be applied in a research project investigating the use of mobile robots in "smart buildings".

Warwick believes his researchers' involvement in robot volleyball could produce benefits in developing ways of sharing experiences between groups of learning robots and in creating "languages" that robots can use to communicate with each other.

The serious side to the sponsorship of computer fun and games was highlighted by IBM's involvement in computer chess. Sports robotics could also encourage innovations in the use and development of intelligent systems.

For example, the Lego "robotics invention" kits used in the robot volleyball competition include a "programmable brick", which is the fruit of long-term Lego sponsorship at the Massachusetts Institute of Technology. The RoboCup'99 legged robot league used Sony's AIBO robot, which looks like a chic version of Dr Who's K9 doggy robot. Fellow RoboCup sponsor Sun Microsystems recently introduced Jini network technology for distributed-control environments, such as that found in robotic team sports.

A realistic perspective on robot sports is summarised by Hallam: "These events make AI and robotics more visible and promote the integration of systems. I am not convinced that they are the best way to make progress in robotics, but they are certainly a fairly good way of doing so - while also having fun!"

www.cyber.rdg.ac.uk/research/CIRG/lego (robot volleyball competition); www.robocup.org (RoboCup)