In applications of HTN planning, repeated problems have arisen from the lack of correspondence between HTN tasks and classical-planning goals. We describe these problems and provide a new Hierarchical Goal Network (HGN) planning formalism that overcomes them. HGN tasks have syntax and semantics analogous to classical planning problems, and this has several benefits: HGN methods can be significantly simpler to write than HTN methods, there is a clear criterion for whether the HGN methods are correct, and classical-planning heuristic functions can be adapted for use in HGN planning.

We define the HGN formalism, illustrate how to prove correctness of HGN methods, provide a planning algorithm called GNP (Goal Network Planner), and present experimental results showing that GNP’s performance compares favorably to that of SHOP2. We provide a planning-graph heuristic for optional use in GNP, and give experimental results showing the kinds of situations in which it helps or hurts GNP’s performance.