Shape classification in rotation manipulation by universal robot hand

Abstract

We propose a method for shape classification in rotation manipulation by a multi-fingered robotic hand. The robotic hand has five fingers equipped with tactile sensors. Each tactile sensor can measure the pressure distribution once every 10(ms). Our proposed classification method consists of the following processes: A kurtosis is calculated from the pressure distribution, and it quantifies shape of the current contact surface. By rotating an object and measuring a time-series pressure distribution, the hand obtains a kurtosis pattern. Finally, a degree of similarity is calculated between the kurtosis pattern and a reference pattern through a dynamic programming (DP) matching scheme. This classifies the contact surface shape using thresholds. Laboratory experiments confirm that this method can classify three objects with a high degree of accuracy.