Representing Plan Revisions

Abstract

Autonomous robots acting in changing and partly unknown environments, such as the delivery robot in the DeliveryWorld, often acquire important information while carrying out their plans. These robots cannot commit in advance to a fixed course of action; they have to be flexible and make critical decisions when they get the necessary information or when they cannot post-pone the decisions any longer. As a consequence, their plans contain conditional, iterative, concurrent, and reactive subplans and are not just inflexible sequences of plan steps. To forestall flaws in the behaviors of robots with such flexible plans, planning systems do not only have to project how the robot will change the world as it executes the plan. They will also have to determine which parts of the plan were (projected to be) executed and why. In addition, they have to make their revisions in plans with a complex internal structure.