If a plan is used for refinement, then a planner accepts it as input and determines a new plan that is hopefully an improvement. The new plan may take more problem aspects into account, or it may simply be more efficient. Refinement may be applied repeatedly, to produce a sequence of improved plans, until the final one is executed. Figure 1.19 shows a refinement approach used in robotics. Consider, for example, moving an indoor mobile robot. The first plan yields a collision-free path through the building. The second plan transforms the route into one that satisfies differential constraints based on wheel motions (recall Figure 1.11). The third plan considers how to move the robot along the path at various speeds while satisfying momentum considerations. The fourth plan incorporates feedback to ensure that the robot stays as close as possible to the planned path in spite of unpredictable behavior. Further elaboration on this approach and its trade-offs appears in Section 14.6.1.
Steven M LaValle 2012-04-20