Timezone: »

 
Solving PDDL Planning Problems with Pretrained Large Language Models
Tom Silver · Varun Hariprasad · Reece Shuttleworth · Nishanth Kumar · Tomás Lozano-Pérez · Leslie Kaelbling
Event URL: https://openreview.net/forum?id=1QMMUB4zfl »

We study few-shot prompting of pretrained large language models (LLMs) towards solving PDDL planning problems. We are interested in two questions: (1) To what extent can LLMs solve PDDL planning problems on their own? (2) How and to what extent can LLMs be used to guide AI planners? Recent work by Valmeekam et al. (2022) presents negative evidence for (1) in the classic blocks world domain. We confirm this finding, but expand the inquiry to 18 domains and find more mixed results with a few clear successes. For (2), we propose a simple mechanism for using good-but-imperfect LLM outputs to aid a heuristic-search planner. We also find that the LLM performance is due not only to syntactic pattern matching, but also to its commonsense understanding of English terms that appear in the PDDL.

Author Information

Tom Silver (MIT)
Varun Hariprasad
Reece Shuttleworth (Computer Science and Artificial Intelligence Laboratory, Electrical Engineering & Computer Science)
Nishanth Kumar (Massachusetts Institute of Technology)

Nishanth Kumar is a Ph.D. student in the LIS Group at MIT CSAIL, where his research is supported by an NSF GRFP fellowship. Nishanth's research interests lie in enabling robots to exhibit long-horizon, multi-task intelligent behavior in real-world scenarios. To this end, his work seeks to synthesize ideas from a number of sub-fields of AI, including Reinforcement Learning, Task and Motion Planning, Program Synthesis and Neurosymbolic AI. Previously, Nishanth obtained a Bachelor of Science in Computer Engineering from Brown University, where he was a Goldwater Scholar, CRA Outstanding Undergrad Researcher Award Finalist, and was named the Outstanding Senior in Computer Engineering upon graduation.

Tomás Lozano-Pérez (Massachusetts Institute of Technology)
Leslie Kaelbling (MIT)

More from the Same Authors