Behavior-Based Planning

In domains such as mobile robot navigation, the dominant characteristics which must be addressed by an intelligent agent are the incomplete and uncertain knowledge of its environment, uncertainty in the current state of a complex system, as well as uncertainty in the effects of the agent’s own actions. In order to function effectively in unstructured, unknown, and dynamic environments, planning systems cannot generate a plan a priori that can be expected to perform reasonably in the face of such uncertainty, nor can they anticipate all contingencies that may arise. Planning systems must be reactive in the sense that their decisions must take into account current information and state at all times, proceeding in a data-driven manner, rather than attempting to impose unrealizable plans in a top-down fashion. Some key issues to be considered in the design of a planning and control architecture are whether the architecture should be centralized or distributed, whether the reasoning should be reactive or deliberative, and whether control should be top-down or bottom-up. In addition, there is a fundamental choice to be made in the method by which information from multiple sources is combined, via sensor fusion or command arbitration. These issues, which are of course interrelated, should not be treated as dichotomies, but rather as continuous spectra to be considered as design tradeoffs to be combined for the desired capabilities of the system. The question then becomes not "which one?" but rather "how much of each?" for a particular class of domains.