The article presents a study on the problem of generating plans for given natural language planning task requests. Large Language Models (LLMs) are great at natural language processing but not so good at planning. Classical planning tools are excellent at planning tasks but require input in a structured language like the Planning Domain Definition Language (PDDL). The researchers propose a Translate-Infer-Compile (TIC) approach, which uses an LLM for generating a logically interpretable intermediate representation of task descriptions, deriving logically dependent information, and generating the target task PDDL. This approach reduces LLM errors and achieves high accuracy in task PDDL generation.
Publication date: 9 Feb 2024
Project Page: https://github.com
Paper: https://arxiv.org/pdf/2402.06608
