PDDL is the de‑facto standard language in AI planning, used to describe planning domains—sets of instances that share the same predicates and action schemas. The language, however, lacks a way to specify the concrete instance set, so constraints on initial states, goal conditions, or a particular subset cannot be expressed directly. Consequently, instance creation has traditionally relied on hand‑crafted generators tailored to each domain or subset. Recent work has explored reasoning‑ or learning‑based automation, but these approaches struggle with scalability. This paper introduces an alternative that leverages large language models (LLMs) to produce instance‑generation programs, coupled with prescribed checks that guarantee soundness. Experiments show that the automatically generated generators can efficiently produce large numbers of valid and diverse instances while providing theoretical correctness guarantees.
Review