Generating Domain Models for Automated Planning from Natural Language Descriptions

The accuracy of automated planning tasks depends on the quality of the PDDL domains models used to represent them; however, manually developing these models is complex and prone to errors. While Large Language Models (LLMs) can process natural language (NL) descriptions of planning tasks, generating accurate and complete PDDL domains remains a significant challenge. In this paper, we propose a modular approach for constructing PDDL domains from NL descriptions through a series of structured extraction and refinement steps involving LLMs. Experiments with expert human evaluators and an LLM-as-a-judge on 30 benchmark planning tasks from the TEXT2WORLD dataset show that our approach generates syntactically correct and semantically coherent domains without the need for structured NL inputs, demonstrating its robustness and generality.