NL to PDDL

Abstract

Automated Planning (AP) is a key component of Artificial General Intelligence and has been successfully employed in applications ranging from scheduling observations of Hubble Space Telescope to generating dialogue agents. A significant bottleneck for its widespread adoption is acquiring accurate domain models which formally encode the planning problem’s environment. Traditionally, these domain models have been hand-coded by human experts and knowledge engineers. However, manually encoding domain models is an increasingly difficult task when one moves away from toy domains towards complex real-world problem scenarios.
To resolve this, the AP community has developed several systems to automatically acquire domain models from valid sequences of actions called plans. This approach has two significant issues. First, the generated domain models might be incomplete, error-prone, and hard to understand and/or modify. Second, most domain learning approaches are based on data-intensive inductive learning, which needs large quantities of structured data (plans) to converge. This data is seldom available without an accurate domain model, which leads to a causality dilemma.
To mitigate these issues, we take advantage of readily available and easy to craft Natural Language (NL) data. We present a pipeline called NLtoPDDL, which takes as input a classical domain’s process manual written in a natural language and outputs its Planning Domain Definition Language (PDDL) model. Specifically, NLtoPDDL does this in two steps: first, it combines pre-trained contextual embeddings with an approach developed in previous research, called EASDRL that extracted structured plans from NL data using Deep Reinforcement Learning (DRL), and a consequence, NLtoPDDL beats the EASDRL model which is the current state-of-the-art on action sequence extraction problem; second, it uses the trained DRL model from the first step to extract structured plans from a domain process manual and employs a modified Learning Object Centered Models (LOCM2) algorithm to one-shot learn a PDDL model. Finally, we showcase the effectiveness of our pipeline on four planning domains of varying complexities, by evaluating our learned domain models for soundness, completeness, validity and intuitiveness.