Unconventional integration of the fundamental elements of soft computing and traditional methods in adaptive robot control

Traditional fuzzy controllers or artificial neural networks are constrained by standardized formal procedures and structural restrictions. While in general it is guaranteed that these approaches or their integration can solve a quite wide class of problems, their “orthodox” application may result in too complicated control not exploiting the peculiarities of the given task under consideration. The aim of the paper is to demonstrate that unconventional integration of simple elements such as classic PID/ST, regression analysis, saturated sigmoid transition functions, fuzzy sets and uniform structures obtained from the Lagrangian classical mechanics instead of the connection structure of a feedforward artificial neural networks can result in a simple and efficient adaptive control for robots involved in unknown environmental dynamic interaction. Due to their simplicity and reduced number of parameters real time tuning can be carried out in these structures. Most of the parameters are independent of the particular problem to be solved and neither “scaling problems” nor “network paralysis” occur during the learning phase. It is concluded that the different components of the control can successfully co-operate in finding the “proper” system model even in the case of very rough initial model estimation and external interaction