Neural Networks Based Interaction Matrix Approximation for IBVS Applications

Abstract

Visually guided robots should interpret what they see and they should accomplish their task by acting according to this interpretation. The method of using visual information as a feedback signal in a closed-loop robot system is named as visual servoing (VS). As the most popular type of VS, image-based visual servoing (IBVS) may encounter a common problem in realization: singularity of the pseudoinverse of the interaction matrix. Although the pseudoinverse of this matrix for IBVS can be applicable, the control law is useless in the case of singularities.In this study, the pseudoinverse of the interaction matrix is approximated with different types of trained neural networks to get rid of the singularity problem. The results of these approximations are given with MSE values. © 2019 IEEE.