CrynodebMultirobot control architectures often require robotic tasks to be well defined before allocation. In complex missions, it is often difficult to decompose an objective into a set of well defined tasks; human operators generate a simplified representation based on experience and estimation. The result is a set of robot roles, which are not best suited to accomplishing those objectives. This thesis presents an alternative approach to generating multirobot control algorithms using task abstraction. By carefully analysing data recorded from similar systems a multidimensional and multilevel representation of the mission can be abstracted, which can be subsequently converted into a robotic controller.
This work, which focuses on the control of a team of robots to play the complex game of football, is divided into three sections: In the first section we investigate the use of spatial structures in team games. Experimental results show that cooperative teams beat groups of individuals when competing for space and that controlling space is important in the game of robot football. In the second section, we generate a multilevel representation of robot football based on spatial structures measured in recorded matches. By differentiating between spatial configurations appearing in desirable and undesirable situations, we can abstract a strategy composed of the more desirable structures. In the third section, five partial strategies are generated, based on the abstracted structures, and a suitable controller is devised. A set of experiments shows the success of the method in reproducing those key structures in a multirobot system. Finally, we compile our methods into a formal architecture for task abstraction and control.
The thesis concludes that generating multirobot control algorithms using task abstraction is appropriate for problems which are complex, weakly-defined, multilevel, dynamic, competitive, unpredictable, and which display emergent properties.