The OxfordRescue Team represents the University of Oxford in RoboCup's Virtual Robots competition.

A search and rescue robot designed for rugged terrain.
Credit: news.bbc.co.uk

RoboCup is an international organization that promotes the development of robotics and artificial intelligence. Universities from around the world take part in various competitions that involve challenges spanning many fields of research. Though RoboCup's initial motivation was robot football (a stated goal is to develop a team of robots that beats the best human team by 2050), research has since branched into a number of other fields. The most notable of these is Search and Rescue, with RoboCup Rescue founded in 2001.

RoboCup Rescue now comprises three competitions:

• The Rescue Robots competition, in which real robots must traverse obstacles and perform a number of tasks demonstrating mobility and navigation
• The Agents Simulation competition, in which intelligent distributed decision-making must minimize disaster in a simulated large-scale scenario (about the size of a city)
• The Virtual Robots competition, in which intelligent exploration methods are developed for teams of robots exploring a simulated small-scale disaster zone (about the size of a city block, or the inside of a building)

A P2AT robot (left) and a Zerg robot (right) in the Yellow Arena of the USARSim simulator used in the Virtual Robots competition.

OxfordRescue participates in the Virtual Robots competition. A number of very interesting problems must be solved for a team to be successful in this league. These include:

• Navigation
• Mapping
• Victim Detection
• Exploration
• Development of a Practical Human-Robot Interface
• Communication
• Cooperation

The Virtual Robots competition has only existed since 2006, so 2008 will be the third time that teams from around the world compete. OxfordRescue has not participated before, so it will be our first time.

Our 2008 team description paper can be found here.

In short, since we are new to the competition, we hope to build on work that has proven successful for others. For example, the University of Amsterdam's UvARescue team has a reputation for excellent mapping, while STEEL, the University of Pittsburgh's team, has developed a highly useful robot-user interface. Combining their excellent ideas would likely be of great benefit.

However, we also hope to move into a new direction by applying learning to some of the problems offered by the competition. In particular, we are interested in improvement of cooperation and communication using learning techniques. Reinforcement learning (especially Q-learning) has proven to be very useful for a number of other RoboCup competitors (e.g. the University of Osnabrueck's Brainstormers, current 2D simulation champions), so it could likely be helpful in solving problems in the Virtual Agents competition.

Recent ideas include representation of the team of robots as a coordination graph, and using Q-learning to learn action payoffs for individual robots. In a communication-limited scenario, robots could learn to take on different roles, and this in turn could be useful to relay information multi-hop through the team. While some robots continue exploring, others could become "Messenger" robots that transmit information through the team.

That said, the team is still in an early stage of development! Currently we are focussing on many low-level tasks, and it may be some time before learning can be properly implemented.

Currently the OxfordRescue team is being developed by Julian de Hoog, under supervision of Dr. Stephen Cameron.

We welcome suggestions and ideas! Please email Julian.