Robotic perception can have simple and effective sensing functions that are unreachable for humans using only the isolated tactile perception method, with the assistance of a triboelectric nanogenerator (TENG). However, the reliability of triboelectric sensors remains a major challenge due to the inherent environmental limitations. Here, an intelligent tactile sensing system that combines a TENG and deep-learning technology is proposed. Using a triboelectric triple tactile sensor array, typical characteristics of each testing material can be maintained stably even under different contact conditions (touch conditions and external environmental conditions) by extracting features from three independent electrical signals as well as the normalized output signals. Furthermore, a convolutional neural network model is integrated, and a high accuracy of 96.62% is achieved in a material identification task. The tactile sensing system is exhibited to an open environment for material identification and the real-time demonstration. Compared to the complex process that humans must integrate multiple sensing (touching and viewing) to accomplish tactile perception, the proposed sensing system shows a huge advantage in cognitive learning for the visually impaired, biomimetic prosthetics, and virtual spaces construction.
Keywords: convolutional neural networks; material identification; open environment; tactile sensing; triboelectric nanogenerators.
© 2022 Wiley-VCH GmbH.