Haptic Adjective Recognition Dataset Construction and Verification

Researchers have been increasingly interested in understanding how humans describe the haptic sensation of objects, aiming to enhance the adaptability of robots in complex tasks. Adjective recognition aims to furnish robots with high-level attribute feature information about their environment. In existing adjective recognition research, the PHAC-2 Dataset has been commonly used for evaluating algorithms. In order to promote adjective recognition algorithm research, we create a new dataset to evaluate adjective properties, named the Haptic Adjective Recognition (HAR) Dataset. This dataset is generated from sliding experiments using a bionic fingertip equipped with force sensors and IMUs. In addition, we propose an adjective recognition method to evaluate the plausibility of the HAR Dataset. We conduct extensive experiments, and the experimental results demonstrate that: (i) our model is highly effective on seen materials, achieving classification accuracy of 98.68% and F1 score of 97.44
^%
; (ii) our model exhibits certain generalization capability on unseen materials, though with some unsatisfactory test results on some specific materials. The results prove the reasonableness of the HAR Dataset distribution, while also underscore the potential of exploring algorithms with robust generalization capabilities for unseen materials. The complete HAR dataset is available for download on Google Drive.