Ph.D. received on: 8/7/1997E-mail: urso@ias.unipa.it
Tutor: Prof. Francesco Alonge, University of Palermo
___________________________________________________________________________________________________________
Neuromorphic Control of Robotic Manipulators ___________________________________________________________________________________________________________Advisor:
Prof. Tommaso Raimondi, University of Palermo
Prof. Francesco Alonge, University of Palermo
Summary:
The Ph.D. thesis deals with the study of adaptive neural controllers for rigid robots. For these dynamic systems model-based control techniques have been extensively used. However, when the mathematical model is not known with a certain accuracy the above model-based control techniques fail to obtain good performances. In this case, it can be convenient to apply control techniques which do not require a priori knowledge of the mathematical model of the system and, consequently, allow us to take into account unmodelled dynamics and structure and parameter variations.
In this thesis it is proposed a control technique based on Artificial Neural Networks (ANN). More precisely, an ANN is implemented as a feedforward controller with the aim of identifying the inverse dynamics of the manipulator. The structure of the ANN, in terms of hidden layers, number of neurones and connections between contiguous layers, is optimized by means of a Genetic Algorithm, using a particular training set suitable for a task-independed learning.
To cope with parameter variations and structural modifications in the mechanical system, the backpropagation algorithm is implemented for on-line adaptation of the Net weights.
Several simulation experiments and experimental tests show that genetic optimization of the Net structure leads to improvements in inverse modelling and capability of generalization, whereas genetic optimization and on-line adaptation give good capabilities to cope with parameter and structural modifications.
_______________________________________