Recurrent Self-Organizing Maps for Recognition of Grasp Sequence of Movements

Research on controllers for robot hands with multiple fingers has emphasized the approach called programming by demonstration. In such an alternative, one aims to develop a robotic system capable of learning grasp tasks through demonstrations of a human operator. Hence, recognition of a particular grasp is often a stage of the learning procedure. We present a recurrent Self-organizing map (SOM) to identify a grasp before it is finished, i.e., during the movement of a robotic hand. In our experiments, the recognition system, simply based on fingers position, obtained 80% of correct recognition of 31 different grasps. The average value of correct categorization of grasp was 87%.