Teaching Large Language Models to Create Task Plans for Robotic Suturing

Megan Zhou, Mahdi Tavakoli

Keywords: Large language models, task and motion planning, robotic surgery, surgical autonomy

Abstract: Surgical robots have the potential to reduce the rate of surgical errors as they can perform tasks with more precision than humans. Autonomous robotic surgery presents challenges, as the complexity of surgical tasks requires high-level decision making and motion execution. We propose a workflow in which we teach a Large Language Model (LLM) to generate machine-readable task descriptions for robotic suturing, which can be integrated into task and motion planning pipelines. We prompted the model with surgical procedure description texts and instructed it to construct Planning Domain Definition Language (PDDL) files for the suturing task, which we then inputted into a task planner to generate a symbolic action sequence. We found that LLMs are capable of producing reliable task plans, although prompt refinement is critical to minimizing syntax and logical errors.

Submission Number: 4