The situated control component of most autonomous agents has traditionally been either a planner or reactive system. However, there is general agreement that planning and reaction represent two ends of a spectrum. These action selection methods differ in, among other things, their treatment of representation. Planners per form projection on an internal world model, while pure reactive systems store no internal representation, using "the world as its own model". We seek to move situated controllers, or perception/action systems, toward the middle of this spectrum. We have developed a perception/action (PA) system which incorporates task dependent representations. These representations do not constitute the full world model of classical planners, but they do provide action selection information beyond the current sensory data used by reactive systems. We will show that our representations are compact and maintainable and that they make our PA system more effective than a stateless reactive system. We will also discuss the implications a PA system using such representation has on the rest of an agent's architecture. The discussion will be carried out in the context of an application involving both planned and dynamic components.
perception/action systems, agent architecture, representation