Title:	NL2Plan: Robust LLM-Driven Planning from Minimal Text Descriptions
DNr:	Berzelius-2024-279
Project Type:	LiU Berzelius
Principal Investigator:	Elliot Gestrin <elliot.gestrin@liu.se>
Affiliation:	Linköpings universitet
Duration:	2024-07-30 – 2024-12-01
Classification:	10205
Homepage:	https://arxiv.org/abs/2405.04215
Keywords:

Abstract

Planning is one of the classical AI disciplines and entails generating a sequence of applicable actions to transform an initial state to a goal state. Powerful classical planners have been developed over decades to solve this problem, but require modeling input tasks in tedious and error-prone formats such as PDDL. In contrast, novel large language model (LLM) based planning methods are easier to apply, but lack guarantees and have been shown to perform poorly. Recent work has begun to combine these approaches, but still requires various degrees of domain adaptation and expert input. We've developed NL2Plan, the first domain-agnostic LLM-driven planning system. It uses an LLM to incrementally analyze the user input before creating a complete PDDL description, which is finally solved by a classical planner. This framework enables the use of classical planners by untrained users, can support expert users in knowledge engineering and preliminary experiments show that it outperforms directly applying an LLM for planning. NL2Plan is currently presented in a peer-reviewed workshop paper with researchers from several universities expressing interest in the project. We aim to publish a finalized version of NL2Plan in a conference paper at the A* conference ICAPS. For this, we will evaluate the NL2Plan framework across 6 domains and several tasks using the 70B parameter Llama-3 LLM and the published NL2Plan code. As follow-up work, we're planning to fine-tune Llama-3.1 for the specific task of generating planning tasks. The project team consists of PhD student Elliot Gestrin, Prof. Marco Kuhlmann and Assoc. Prof. Jendrik Seipp, all from IDA at LiU.