Can We Automatically Label Expressive Robot Arm Motion Qualities? Open Dataset and Machine Learning Results

Abstract

Robot motion impacts onlooker perception of robots' social and functional goals, whether intentionally or unintentionally. Neighboring fields such as dance can help roboticists understand the impacts of motion using tools like the Laban efforts, one of the primary modes of annotating and interpreting motion expressivity, but many past efforts at the intersection of robotics and Laban movement analysis feature unscalable hand-designed motions or offer limited information about factors that influence each individual Laban effort. We sought to begin addressing these gaps by collecting, analyzing, and open-sourcing an expressive motion dataset comprising robot arm motion demonstrations (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$N = 30$</tex-math></inline-formula> demonstrators, with 219 total viable demonstrations) that span the Laban effort axes. We demonstrate an 82.0-92.3% accuracy for expressive motion label classification (ranging slightly across expressive axes, approaches with vs. without feature reduction and approaches with sequential data). We propose that the most informative features may guide future methods for automatically generating expressive motion. Individuals with interest in robot expressive motion can benefit from the dataset itself, as well as the machine learning results.