Contact states recognition in robotic part mating based on support vector machines

Abstract

In order to autonomously execute part mating, an intelligent robot should be able to carry out active compliant motion. Contact states recognition machine is a missing link for implementation of this kind of motion within an intelligent robotic assembly system. In this paper, we present an approach to design contact states recognition machines for various types of part mating problems. We have chosen generalized mating force as a basic feature measured from the process. Starting from force measurements, online recognition is carried out using class boundaries and transduction mappings obtained during offline training. The basis for the proposed offline training procedure is not experimental data but a mechanical model of the part mating process. This enables supervised training without requiring numerous experiments. Furthermore, this has allowed extraction of qualitative features from the analytical model of mating force. To provide good time localization and phase correctness, we have utilized discrete wavelet transform for feature extraction. The obtained patterns have been classified using support vector machines to obtain a recognition machine with good generalization properties. The proposed machine is elaborated and experimentally verified using case study of cylindrical part mating with chamfer crossing. We have used quasi-static model of insertion force as a starting point in the training process. Exploiting some characteristics of Daubechies wavelets, we have managed to extract features that are independent on characteristics of the concrete cylindrical part mating. Finally, the generated machine was evaluated using intensive real-world experiments. Herein we have shown that performance of the generated contact states recognition machine was excellent.