Compiling Expressive Planning with Data Types

Abstract

Planning involves selecting action sequences to achieve goals from initial conditions, a fundamental task arising in contexts ranging from daily tasks to industrial processes. The Planning Domain Definition Language (PDDL) is the standard formalism for describing such problems, yet previous work has highlighted some of its limitations in expressivity. In particular, many real-world planning problems inherently exhibit structural patterns such as sets, arrays, and counting. These are features of the problem domain itself that are cumbersome to model directly in PDDL. We present a high-level modelling language that compiles to PDDL, supporting richer data structures and operations including arrays, sets, Boolean counting expressions, integer range variables, and bounded integer parameters in actions. This allows users to express problems in terms closer to their natural structure, without committing to a particular low-level encoding. From a single high-level specification, our framework can automatically generate and explore multiple PDDL encodings with different performance trade-offs in solving time, memory usage, and plan quality. We demonstrate our approach on a set of benchmark problems involving complex structural patterns. Our results show that compiled models are competitive with established hand-crafted models from the literature, and can sometimes outperform them.