Learning manipulative skills with ART

Abstract

The research reported in this paper is related to the creation of self-adapting robots that are capable of learning manipulative skills online. The investigation includes the design of a novel neural network controller (NNC), which is based on the adaptive resonance theory (ART) and a dynamic knowledge base, whose knowledge is regulated by specific assembly operations. A force/torque (F/T) sensor was attached to the robot's wrist and this was the only information available to the NNC during the assembly operations, since the precise location of the components was unknown. The knowledge is enhanced online, based on the success in predicting the motion that reduces the constraint forces. The results demonstrate the generalisation capability of the NNC by learning the assembly of different part geometries using the same initial knowledge base. The learning time for a complete new operation was achieved in approximately 1 minute.