Learning Refinement Cost Estimators for Bilevel Planning

Abstract

Bilevel planning is an effective approach for solving complex task and motion planning (TAMP) problems with continuous state and action spaces, that involves first searching for a high-level abstract plan and then refining it into a sequence of lowlevel actions. Although the low-level refinement process is a significant contributor to the total time needed to solve a task, this cost is typically unaccounted for during high-level planning. This can result in undesirable behavior if abstract plans that are difficult or even impossible to refine are selected over alternatives that may be slightly longer but can also be refined significantly faster. This work develops a method for learning to estimate the cost of refining an abstract plan and a framework for using the estimator to guide high-level search in a bilevel planner. We demonstrate using two environments that our proposed approach considerably improves on the combined planning and execution cost required for tasks compared to several baselines, including a standard benchmark bilevel planner and alternative estimator models.