Unsupervised surface roughness discrimination based on bio-inspired artificial fingertip

Different from texture classification, surface roughness discrimination is more challenging in the development of tactile sensing because of limited discriminative information. In recent years, it is receiving more and more attentions from researchers in various fields, most of which are based on supervised learning. But frequently all we have is unlabeled dataset with very limited prior information, i.e., labels are not available to train the discrimination models. Lacking the 'teaching' process, it becomes rather difficult to locate the boundary of different classes. In this paper, the ability of unsupervised surface roughness discrimination is explored based on our developed bio-inspired artificial fingertip. At first, the original signals are analyzed and discriminated with the most widely used unsupervised algorithm (K-means clustering). Then the technique of discrete wavelet transform and algorithm of sequential forward selection are utilized to identify the most discriminative features. The unsupervised discrimination results (K-means clustering) are presented and compared based on different distances including Squared Euclidean, Cityblock, and Cosine. The highest test accuracy reaches 72.93% ± 12.55% when the distance of Squeared Euclidean is adopted with six discriminative features. Finally, another popular unsupervised algorithm, self-organizing maps neural network that is different from clustering, is also applied in discriminating surface roughness with lower accuracy. The results show that unsupervised learning algorithms with our developed tactile fingertip are capable to discriminate surface roughness, which have great potentials in robotics and autonomous applications.