Experimental investigation of surface identification ability of a low-profile fabric tactile sensor

Humans usually distinguish objects by sliding their fingertips on the surface to feel the texture via mechanoreceptor underneath the skin. We have developed a human-imitated system for robotic fingertip to sense object's texture via sliding action. Design of the sensory skin was inspired by the localized displacement phenomenon of a sliding soft fingertip ([1]) to capture stick-slip events on the contact surface that mainly represent texture characteristics. The soft skin is knitted by electro-conductive tension-sensitive yarns, then covered over a hemispherical fingertip. The pile-shaped surface of the fabric sensor enhances tangential traction detection ability of the sensor, even though the normal load is also sensible. Our aim is to exploit this sensor in applications regarding relative sliding between the touched object and the surface of the sensor, such as slip detection ([2]), and surface identification in this paper. In surface encoding, we have experimentally investigated ability of the fabric sensor in recognition touched objects via multiple machine learning algorithms, such as naive Bayes, Multi-Layer Artificial Neural Network (ANN) with input extracted from autoregressive models, and ANN with input extracted from Discrete Wavelet Transformation (DWT), have been trained to distinguish three typical textures. As a result, we have found that the last method outperforms the remains with an average successful rate of 90%.