Leveraging Multi-label Correlation for Tactile Adjective Recognition

Tactile sensing is a complementary mode to visual and auditory perception, which plays a vital role in autonomous robotics. Improving the tactile sensing capability of robots with the aid of machine learning methods has attracted increasing attention. Various approaches have been developed for the understanding of the haptic adjectives. However, most of these methods use very complex features, and cannot exploit the correlation fully among the multiple tactile adjectives. An object is often described by more than one tactile adjective. On this basis, the tactile understanding of multiple adjectives can be formulated as a multi-label classification problem. To solve this problem, we exploit the potential relation among different adjectives and test the effect of the label correlation with different tactile classifiers. We design simpler tactile features and use four methods, standard support vector machine (SVM), k-nearest neighbors (KNN), and ranking support vector machine(RANK-SVM) to classify the tactile adjectives. Finally, extensive experiments are performed on the Penn Haptic Adjective Corpus 2 dataset, and the experiment results show that the proposed methods can achieve a higher classification F-1 score than the competing methods.