Exciting news—my workshop paper, “Creating PDDL Models from JavaScript Using LLMs: Preliminary Results,” has been accepted into AAAI 2025! This work stems from research I did during my time at the Naval Research Lab this past summer. I’m thrilled to be heading to Philadelphia to present it and while there to track down a Philly cheesesteak (or two).
In AI planning, automated planners require formal domain models to generate correct plans. But creating these models isn’t always straightforward—as many know, writing them can be time-consuming and often requires specialized domain knowledge.
In this work, we tackle this challenge by creating a two-stage pipeline that combines natural language and operational models to generate these planning models. Specifically, we convert JavaScript code from an existing robotic controller (in this case, corresponding to Minecraft actions) into PDDL, not only extracting a valid planning model but also testing the the limits of what large language models (LLMs) can do. Based on our findings, we highlight a few techniques and other research directions we feel could enhance the effectiveness of LLMs in future AI planning endeavors.
Check out the paper here.