Go to AAAI 2025 Workshop LM4Plan homepageCreating PDDL Models from Javascript using LLMs: Preliminary Results
Keywords: Large Language Models, Natural Language Processing, Knowledge representation and engineering, robotics
Abstract: Systems that act within an environment (e.g., robots, game agents) often have at least two layers of planning: a descriptive task-planning layer that describes what to do and an operational acting layer that states how to accomplish a task or action. Because they provide platform-specific details, operational models present a rich source for constructing task planning models. While LLMs have been used for generating task models from natural language, less work has examined how to incorporate operational models. We develop a pipeline that produces PDDL task models by combining natural language and operational JavaScript models for a Minecraft agent. We evaluate the pipeline's ability to produce correct actions under varying conditions. Our results show that the pipeline sometimes generates valid actions, although their correctness fluctuates depending on the input and parameters. We provide six points for consideration of future work in this area.
Submission Number: 31
Loading