Connecting planning and acting via object-specific reasoning

Instructions from a high-level planner are in general too abstract for a behavioral simulator to execute. In this dissertation I describe an intermediate reasoning system--the OBJECT-SPECIFIC REASONER--which bridges the gap between high-level task-actions and action directives of a behavioral system. It decomposes task-actions and derives parameter values for each action directive, thus enabling existing high-level planners to instruct synthetic agents with the same task-action commands that they currently produce. The OBJECT-SPECIFIC REASONER\u27S architecture follows directly from the hypothesis that action representations are underspecified descriptions, and that objects in the same functional category are manipulated in similar ways. The action representation and the object representation are combined to complete the action interpretation, thereby grounding plans in action. The OBJECT-SPECIFIC REASONER provides evidence that a small number of object functional categories, organized in a taxonomy, makes possible a simple and elegant reasoning system which converts task-actions to action directives. To test the theory behind the OBJECT-SPECIFIC REASONER, I applied the implementation to three different 3D graphical domains