The combination of recent improvements in robotic tactile sensors and the development of new neural network paradigms and tools currently allows to investigate of tactile perception problems in robotics by mimicking last findings in neurophysiology. In this paper we describe an antropomorhic approach to robotic tactile perception, which is characterized by a strong biological inspiration both in the hardware tools and in the processing methodologies. As a first step toward this goal, an antropomorphic robotic finger including three different kinds of tactile receptors (static, dynamic and thermal) has been designed and fabricated, and the problem of sensory-motor control for feature enhancement micro-movements has been investigated. A neural network architecture is proposed for what we believe to be the basic problem of tactile perception, that is the autonomous learning of sensory-motor coordination. Experiments are described, system performances are analyzed and possible applications are outlined.