Date of Award

Winter 2023

Project Type

Thesis

Program or Major

Computer Science

Degree Name

Master of Science

First Advisor

Samuel Carton

Second Advisor

Samuel Carton

Third Advisor

Momotaz Begum

Abstract

A common scenario in robotics is needing to change the behavior of a deployed robot according to new requirements. This usually requires intervention from the creators of the robot, which both labor- and time-intensive. In this work, we explore the capability of a large language model (LLM) to address this issue by modifying the existing behavior of robots represented using PDDL. Specifically, we prompt the GPT-4 LLM to generate new PDDL problem files given an existing problem file and a natural language description of how the robot's behavior should change. We evaluate both zero- and one-shot versions of this method on a modified version of an existing PDDL dataset, and compare our results with a baseline approach. We find that this problem is challenging for LLM to solve and discuss possibilities for improving performance.

Share

COinS