A neural network-based system for tactile exploratory tasks

This paper deals with the problem of object identification by touch. In experiments carried out on human subjects, five tactile primitives were identified, controlling the exploratory strategy of the fingertip on the object. These primitives have been used for setting up the algorithms driving a robotic gripper during the exploration of an unknown object. An artificial tactile system has been designed for the recognition of prism shaped rigid objects. It is based on artificial tactile sensors attached on each fingertip of a two finger robot gripper, giving a sampled image of the normal stress distribution on the sensor´s surface. The stress images are processed by a three-layer MLP which recognizes the tactile primitive and gives its direction on the sensor surface. The basic strategy consists of reaching an edge of a face and following all the edges of the same face. At the end of a face exploration, an ART neural network classifies the face touched. When all the faces have been explored, a second ART network extracts the shape of the touched prism, by combining the trajectories accomplished by the grip with the characteristics of faces