NLtoPDDL

Abstract

Existing automated domain acquisition approaches require large amounts of structured data in the form of plans or plan traces to converge. Further, automatically-generated domain models can be incomplete, error-prone, and hard to understand or modify. To mitigate these issues, we take advantage of readily-available natural language data: existing process manuals. We present a domain-authoring pipeline called NLtoPDDL, which takes as input a plan written in natural language and outputs a corresponding PDDL model. We employ a two-stage approach: stage one advances the state-of-the-art in action sequence extraction by utilizing transfer learning via pre-trained contextual language models (BERT and ELMo). Stage two employs an interactive modification of an object-centric algorithm which keeps human-in-the-loop to one-shot learn a PDDL model from the extracted plan. We show that NLtoPDDL is an effective and flexible domain-authoring tool by using it to learn five real-world planning domains of varying complexities and evaluating them for their completeness, soundness and quality.