Abstract
Because of the realities of converting commands into motion there is always a degree of uncertainty as to how accurately an intended action has been executed. This situation has received considerable attention in the study of robot navigation, and the common solution is to calculate the degree of uncertainty in position resulting from a chain of commands and to take measures to reduce this uncertainty when it exceeds some limit. These calculations necessarily make assumptions about the factors which introduce uncertainty, such as wheel slippage per foot or backlash. It is difficult, however, to foresee what all these factors are and even more difficult, in some cases, to develop good models for them. This paper addresses the problem from a different perspective. Rather than calculating the degree of positional uncertainty, it is controlled by what might be called "perceptual servoing", accomplished using a combination of low level tracking and high level perceptual verification of "milestones", sensory events the robot should perceive during the execution of a plan. Perception, planning and execution are integrated in this system: the planning activities are able to reason about what should be perceived at various stages so that it can generate milestones; and execution activities use perception to continuously monitor progress. After a brief exposition of a new robot project, of which this effort is a part, this paper describes a hierarchical representation of space into locales or space packets. This representation is designed to facilitate navigational planning, reasoning about perceptual milestones, and perception of the milestones themselves. Following this discussion, the planning, execution and monitoring activities are described. Planning is done hierarchically, taking advantage of the hierarchical representation of space, and incrementally, expanding only the early parts of the trip in an effort to keep memory requirements and replanning computation under control. Execution is done in small steps using low level perception to monitor progress. Finally, plan monitoring activities construct operators from descriptions in the model and use milestones to verify that the overall physical motion is conforming to plan. The research project described here is quite new and results are just becoming available. Consequently this paper focuses on the design of what is a very top down system which integrates perception and reasoning activities, striving towards continuous and plan-directed perception, used here to control the amount of uncertainty.
