Forward Chaining Hierarchical Partial-Order Planning

Andrew Messing, Seth Hutchinson

Abstract: In recent years, the surge in interest in integrated Task and Motion Planning (TAMP) has caused a resurgence in Task Planning research in the robotics community. Meanwhile, in the broader AI community, combining partial-order planning with grounded forward chaining has become popular for temporal task planning, preserving the ability of partial-order planners to deal with concurrent actions while exploiting the speed of modern-day grounded forward search. In this paper, we present two new planning algorithms. The first, FCPOP, combines full utilization of the delayed action ordering commitment of partial-order planning with grounded forward search guided by a temporally informed heuristic. This results in a planner that is flexible in how it structures plans with respect to action parallelism, creates high-quality plans with low makespans, and computes plans quickly. FCPOP is shown empirically to outperform state-of-the-art temporal planners on several benchmark planning problems. The second planning algorithm FCHPOP introduces hierarchical information in the form of abstract actions. This reduces the number of nodes that need to be explored and speeds up the planning process while still generating quality plans.