An architecture for planning with external information points in a real-time system

Abstract

Recently several AI researchers [MHA +95, HA90, GI89] have acknowledged the limitations of traditional planning methods in real-time systems. Some have addressed the problem by planning within the allotted time and sending only partial results to a real-time scheduling system [GI89, IG90, BD89]. Others have reduced the system's ability to plan in order to meet the time constraints of the system [MHA+95, HA90, PB91]. However, in systems which are not extremely time-critical (i.e. not medical trauma systems), we can sketch a high-level plan at the beginning of the task and interleave the planning of the low-level details during execution. In this paper, we describe a new architecture for robot planning which utilizes domain knowledge stored externally to the system and show how access of this knowledge can be accomplished through the normal task planning mechanism. Because of the structure of this knowledge and our hierarchical planning mechanism, we are able to reduce the complexity of the planning process for real-time applications. The reduced computation also allows for less expensive hardware implementations.