Equipe EEA - Groupe TROP

Three neural structures dedicated to continuous function approximation are evaluated and compared with a view to the implementation of an adaptive kinematic approach for vision-based robotic control. This approach is using an iterative method of nonlinear mapping inversion, developed by S. Lee and R.M. Kil, and based on the pseudo-inverse of the gradient of a Lyapunov function. This method assumes the knowledge of the direct mapping that we aim in our context to estimate by learning. To that end three neural structures are considered: the well-known Ridge Polynomial Networks (RPN) and Radial-Basis Functions networks (RBF), and the contributed Hyperplan based Self-Organizing Maps (HYPSOM). The application of theses structures to robotic control is illustrated in the framework of the position control of a two joint axis robot manipulator. The optimization algorithms used for learning are of the usual types : Least Mean Squares (LMS) and Total Least Squares (TLS).