Aproximação da cinemática inversa de um robô manipulador didático através de algoritmos de aprendizado de máquina

Abstract

The calculation of inverse kinematics is a challenge in robotics, especially for robotic manipulators, and traditional methods are often ineffective. This work proposed the use of machine learning algorithms, focusing on local and global model approaches, to approximate the inverse kinematics of an educational robotic manipulator. Specifically, Multilayer Perceptron neural networks and the K-means algorithm were employed for the local models to segment the operational space. Initially, direct kinematics modeling is carried out using Denavit-Hartenberg notation to generate a database containing information about the Cartesian space and the manipulator’s joint space. This database is then filtered, removing unreachable samples and samples that are close to each other in Cartesian space. The hyperparameter optimization phase for the global models involved adjusting the best architecture of the neural networks, while for the local models, it included determining the optimal number k of clusters that best divided the operational space. The performance evaluation of the algorithms is based on R2 and MSE measures, both in joint space and in Cartesian space, and the results obtained demonstrate that both approaches proved viable for the task at hand, with the local approach outperforming the global one.