Plan Abstraction Based on Operator Generalization

We describe a planning system which automatically creates abstract operators while organizing a given set of primitive operators into a taxonomic hierarchy. At the same time, the system creates categories of abstract object types which allow abstract operators to apply to broad classes of functionally similar objects. After the system has found a plan to achieve a particular goal, it replaces each primitive operator in the plan with one of its ancestors from the operator taxonomy. The resulting abstract plan is incorporated into the operator hierarchy as a new abstract operator, an abstract-macro. The next time the planner is faced with a similar task, it can specialize the abstract-macro into a suitable plan by again using the operator taxonomy, this time replacing the abstract operators with appropriate descendants.