Nova abordagem de aproximação de modelos baseada em aprendizado de máquina aplicada à um manipulador robótico

Abstract

Algorithms based on Computational Intelligence are being applied in robotics to solve modeling problems. In addition, machine learning algorithms are also being indicated in a wide variety of systems applications. In this thesis, using machine learning algorithms, we propose a new model approximation algorithm for a cylindrical robotic manipulator. Initially, the mathematical modeling of manipulators is described, taking into account the kinematic and dynamic models. Also some regression algorithms present in the literature are approached and applied to the system, as well as the proposed algorithm based on committee machines was also applied. This algorithm is based on the support vector regression (SVR) method, where a committee chooses one of three types of SVR structures with kernels: gaussian, polynomial and sigmoidal. The selection is made by an error-based criterion. Therefore, the mathematical modeling of manipulators is described, taking into account the kinematic and dynamic models which are also obtained for the cylindrical manipulator driven by three-phase induction motors mounted on the GPAR. A comparison between the various methods presented is performed using simulation and experimental data obtained from the GPAR cylindrical manipulator, and then cross-validation is performed for all methods. This comparison is performed through tasks performed by the manipulator with constant and variable loads. Tasks with variable load have the objective of simulating possible problems in products during their execution. Finally, the results are compared by evaluation metrics such as root mean square error (RMSE) and adjustable coefficient of determination (R2 a), which confirm that the proposed method presented the best results in most scenarios.