An adaptive learning approach to control contact force in assembly

Abstract

Robotic assembly operations can be performed by specifying an exact model of the operation. However, the uncertainties involved during assembly make it difficult to conceive such a model. In these cases, the use of a connectionist model can be advantageous. In this paper the design of a neural network controller (NNC) based on unsupervised learning is presented. The NNC consists of two stages, adaptation and decision. The first stage based on adaptive resonance theory (ART) classifies and recognises all the contact force patterns, whereas the other stage selects the appropriate arm motion direction. Initial results on the implementation of the NNC, using a 6-DOF PUMA robot with a wrist force/torque (F/T) sensor, demonstrate its ability to learn new or novel contact force patterns fast. If previously learned force patterns are encountered these are accessed directly otherwise memory space is allocated to them without forgetting past events, hence creating a stable system.