A Planning System Utilizing Meta Knowledge

Abstract

A new technique for the robot planning, by which the action sequence of a robot is automatically produced, is proposed. In a conventional system,if the rules for actions of a robot and the state descriptions for a circumstance are given, then a sequence of actions necessary for achieving the goal state is obtained according to a deduction algorithm. In our system, however, the meta knowledge for determining the order of performing subgoals, of which the goal is composed, is automatically deduced at the first stage and then the planning proceeds by utilizing this meta knowledge. Moreover, the system is easy to understand and the planning algorithm is very simple, since the state is described on the basis of the object-oriented concept. The meta knowledge is derived by using a state transition diagram representing whether two subgoals arbitrarily selected are valid or not. The concepts of COMPLETE and quasi COMPLETE are also utilized in the planning system. These concepts are used in order to pick up the one to be handled from objects which are partially ordered by applying the meta knowledge. Improved STRIPS is used in the main part of the planning system. Some examples are shown and this system is clarified to be efficient by comparing with a conventional planning system in computing time.