Abstract
Considers a theory of planning, sensing and action in which the fundamental building blocks are, in effect, the 'recognizable sets'-that is, the places in the world that the robot can recognize and distinguish between. To this end the authors observe that viewing the world through sensors, partitions the world into 'perceptual equivalence classes.' The more information the sensors provide, the 'finer' this partition is. Various possible partitions of the world fit into a lattice structure. This lattice structure captures the information or knowledge state about the world. Using this theory the authors develop notions of task-directed planning and show how history can be used to direct actions and gain information. The authors investigate the structure of the recognizability lattice and its relationship to the world. The authors consider how to compute perceptual equivalence classes both from a model, and incrementally, from the world. Finally, the authors investigate mathematical properties that can help the robot generate disambiguating strategies to gain information about the world to accomplish a task.

This publication has 19 references indexed in Scilit: