Templated vs. Generative: Explaining Robot Failures

Abstract

The need for robots to explain their failures grows as the variety and number of robots deployed in public, homes, and work environments increases. This paper extends our prior work utilizing explanation templates by comparing those Templated explanations to Generative explanations created by a Large Language Model. Our study surprisingly reveals that Templated explanations result in similar or higher perceived intelligence and trust while also being more understandable. Through our findings, we aim to provide considerations for effective robot explanation systems, ultimately enabling people to be able to understand and provide assistance to robots that have encountered unforeseen circumstances.