Roughness discrimination with bio-inspired tactile sensor manually sliding on polished surfaces

As an important application of tactile sensing, the capability of texture discrimination is desired in many researches, such as artificial limbs, dexterous manipulation, and humanoid haptic mechanism, etc. In this research, a bio-inspired tactile sensor containing two perpendicular sensing films was developed and manually controlled to slide across 15 polished surfaces with different roughness. Algorithms of discrete wavelet transform (DWT), sequential feature selection (SFS) and extreme learning machine (ELM) were combined together to form a signal processing system for signal decomposition, feature selection and roughness discrimination respectively. Factors affecting roughness discrimination accuracy were evaluated and compared with human tactile sensing capability. Particularly, influences of starting point and sliding duration on discrimination accuracy were analyzed to identify the optimal signal period for the following analysis. Effects of sampling rate on the discrimination accuracy were then investigated using the ELM classifier. Another two typical classification models, k-nearest-neighbor (kNN) and support vector machine (SVM), of different parameter configurations were compared with ELM in surface roughness discrimination. Results showed that the developed method based on manual sliding of the developed tactile sensor is capable of providing a surface roughness discrimination accuracy of 72.93 ± 10.48% with the real polished surfaces.