Evaluation of tactile feature extraction for interactive object recognition

Abstract

Tactile sensing stands to improve the manipulation and perception skills of autonomous robots. Object and material recognition stand as two important tasks, where tactile sensing can aid robotics. While much work has been done on showing the applicability of specific sensors to recognition tasks, a comprehensive examination of the features used has not been performed. In this paper we thoroughly examine the different components of performing interactive object recognition with tactile sensing. We use a state-of-the-art multimodal tactile sensor, allowing us to compare features previously presented for a number of different platforms. We examine the statistical features, robot motions, and classification approaches used for performing object and material recognition. We show that by combining simple statistical features captured from five robot motions our robot can reliably differentiate between a diverse set of 49 objects with an average classification accuracy of 97.6 ± 2.12%.