Answers are grounded in this paper and cite their sources.
Automatically generating complete and reliable Planning Domain Definition Language (PDDL) models from natural-language task requirements remains challenging because large language models (LLMs) may produce incomplete domain knowledge, inconsistent symbolic definitions, and models that cannot support valid planning. This paper proposes OntoPDDL, an ontology-guided method for automatically constructing complete PDDL Domain–Problem models. OntoPDDL uses a domain ontology to provide explicit knowledge constraints and organizes Domain construction according to the structural dependencies among PDDL components, followed by Problem instantiation within the generated Domain. Layered validation checks formal correctness and domain-semantic consistency, while validation and planner feedback guide targeted repairs to improve planning usability. Experiments on 24 earthquake-rescue tasks show that OntoPDDL achieves final Domain and Problem validation success rates of 100.0% and 91.7%, respectively, and increases planning success from 58.3% to 70.8% after planner-feedback repair, compared with 33.3% for LLM+P. Ablation results further support the roles of ontology constraints, dependency-aware generation, and feedback-based repair under the evaluated configurations. A separate task-semantic evaluation based on information correctness and rule-based information completion yields fidelity rates of 45.8% (11/24) for OntoPDDL and 8.3% (2/24) for LLM+P, indicating that OntoPDDL achieves higher task-semantic fidelity under the evaluated configurations. These results highlight improvements in model reliability and planning usability, alongside remaining challenges in task-semantic fidelity.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: