Explainable SayCan

The Explainability of Service Robots utilizing Large Language Models

Abstract

The proposed system, SayCan developed by Google, employs large language models to interpret ambiguous language instructions and execute a variety of tasks. Specifically, SayCan decomposes a task into a series of subtasks, thereby facilitating the provision of explanatory answers to "what" questions. Nevertheless, it is crucial for an autonomous robot to be able to answer "why" questions, and SayCan does not currently account for such explainability. As such, this study examines SayCan-based explanations for service robots within the framework proposed by the authors. Our findings demonstrate that it is possible to classify explanatory properties in SayCan into several categories. Furthermore, we propose an algorithm that utilizes this classification and large language models to explain in natural language. We validate the effectiveness of our proposed method through implementation in a simulator for the tabletop task.