A data-driven robotic tactile material recognition system based on electrode array bionic finger sensors

In order to solve the problem of lacking interactive information such as hardness, temperature, etc. when the surface attributes of objects are visually perceived, a recognition system based on the tactile perception of bionic finger is proposed. First, an experimental platform of bionic finger tactile signal acquisition system is built, which can simultaneously collect 14 channels of tactile electrode signals. Second, the multi-channel tactile signals are fused based on three methods: correlation function weighted fusion method, horizontal fusion method, and horizontal fusion method according to the weight value for each channel. Then the time and frequency domain signals are obtained from the transformation of fused tactile sequences. Third, a dual-stream SE-CNN network (Squeeze-and-Excitation Block, Convolutional Neural Network) of multi-channel signal fusion is established to extract the time and frequency domain features of the fused tactile data-streams. The recognition accuracy of the dual-stream SE-CNN model can reach 99.375% for 20 kinds of materials. The experimental results show that comparing with the CNN models, which only extract features in the time or frequency domain, the proposed model has a higher recognition accuracy. It can successfully recognize the real material of touched objects and lays a solid foundation for exploration of interactive robot tasks.