Construction of Evaluation Datasets for Deriving Equations of Motion Using Large Language Models

Abstract

In recent years, large language models (LLMs) have garnered significant attention for generating robot motions. However, the safety evaluation of these generated motions has remained superficial, and concerns have been raised regarding the lack of a rigorous dynamical foundation. In response, this study introduces a novel process that derives equations of motion from natural language and images, thereby enabling a physically grounded interpretation of the generated behaviors. Moreover, to quantitatively assess the LLM's comprehension of equations of motion, we constructed a QA benchmark dataset comprising images and their corresponding equations of motion collected from publicly available websites and books. Our evaluation experiments demonstrate that, for certain tasks, the proposed method yields accurate derivations.