A General Methodology for Robotic Haptic Recognition of 3-D Objects

Abstract

Three-dimensional object recognition is a fundamental prerequisite to build versatile robotic systems. This paper describes an approach to the recognition problem that exploits tactile sensing, which can be conveniently integrated into an advanced robotic end-effector. The adopted design methodology is based on the training and classification activities typical of the unsupervised Kohonen neural networks, with a learning phase of the geometric properties of the objects, followed by the operative phase of actual recognition in which the robot explores with its end-effector the objects, correlating the sensorial data with the preceding perceptive experiences. The validity of the novel approach pursued for the design of the haptic recognition system has been ascertained with reference to a high-dexterity 3-finger, 11-degree of freedom robotic hand (the University of Bologna hand), but the underlying methodological issues can be specialized to any robotic dexterous end-effector. The developed prototype system, even though currently referring to a simulated environment, has already shown a satisfactory operative level in recognizing objects belonging to a set of significant cardinality, independently of their pose in the working space.