Deadlines, Travel Time, and Robot Problem Solving

1 Abstract This paper describes some extensions to the reductionist planning paradigm typified by Sacerdoti's NOAH program. Certain inadequacies of the partial ordering scheme used in NOAH are pointed out and a new architecture is detailed which circumvents these problems. An example from the semi-automated factory domain is used to illustrate features of the new planner. Techniques for eliminating unnecessary travel time by the robot and avoiding backtracking due to deadline failures are discussed and their incorporation in the planner is described. 2 Introduction Most planners since NOAH [8] have represented a plan for a set of tasks as a partial (ie., non-linear) ordering of the steps required for carrying out those tasks. NOAH and its successors (eg., NONLIN [9] and DEVISER [10]) employ a partially ordered network of tasks to avoid early and unnecessary commitment to task orderings. The motivation for this is to eliminate backtrack-ing. However, maintaining a consistent partial order is difficult in domains where the fact that tasks actually take time plays an important role: domains in which deadlines or robot travel time are serious considerations. That a partial order leads unavoidably to a deadline failure usually cannot be discovered until an attempt is made to linearize the ordering. However, failure to notice a deadline violation early will require backtracking later. In addition, planning with a partial order is not well suited to efficiently managing factors like travel or machine running time. It's not difficult to represent that moving from one workstation to another takes time proportional to the distance separating the two workstations. However, generating a plan that eliminates unnecessary travel between workstations requires exploring some of the linearized task orderings. It is not until tasks are completely ordered that the source and destination workstations of each movement can be known and the travel time computed with any accuracy. In this paper we describe an approach to planning that combines the use of a partial order with a method for exploring the possible repercussions of that partial order. This approach has been implemented in the FORBIN planner (First Order RoBot INtender). FORBIN is a planner capable of solving a significantly wider class of problems than any of its predecessors. 2.1 The Factory Domain One problem domain that has been used for exploring our approach to planning is what we refer to as the semi-automated factory. In this domain a mobile robot operator wanders about the …