Towards Practical and Grounded Knowledge Representation Systems for Autonomous Household Robots

摘要

Abstract — Mobile household robots need much knowledge about objects, places and actions when performing more and more complex tasks. They must be able to recognize objects, know what they are and how they can be used. This knowledge can often be specified more easily in terms of actionrelated concepts than by giving declarative descriptions of the appearance of objects. Defining chairs as objects to sit on, for instance, is much more natural than describing how chairs in general look like. Having grounded symbolic models of its actions and related concepts allows the robot to reason about its activities and improve its problem solving performance. In order to use action-related concepts, the robot must be able to find them in its environment. We present a practical approach to robot knowledge representation that combines description logics knowledge bases with data mining and (self-) observation modules. The robot collects experiences while executing actions and uses them to learn models and aspects of action-related concepts grounded in its perception and action system. We demonstrate our approach by learning places that are involved in mobile robot manipulation actions. I.