Artificial Neural Networks for the Forward Kinematics of a SCARA Manipulator: A Comparative Study with Two Datasets

Abstract

In the field of robotics and automation, the precise determination of manipulator positions is a fundamental aspect with widespread applications. Despite the significance of analytical methods, solving the Forward Kinematics (FK) for Selective Compliance Assembly Robot Arm (SCARA) manipulator robots can be complex and computationally intensive. Our research introduces a method to address the FK problem for the four degrees of freedom (4-DoF) SCARA manipulator through Artificial Neural Networks (ANNs). We utilize two different generated datasets, one with a fixed step size and another one based on a sinusoidal signal, for training the ANNs. This choice of datasets allows us to assess the ANN s' generalization capabilities across a range of operating conditions. Thereafter, the study is planned to investigate the ANN's performance while employing three distinct training algorithms: Levenberg-Marquardt (LM), Bayesian Regularization (BR), and Scaled Conjugate Gradient (SCG). Through a comprehensive comparison of the performance of different ANN models, different training algorithms, and the two adopted generated datasets, this study provides valuable insights into selecting the most optimal training configurations for the SCARA robot's FK solutions. We employ the Mean Squared Error (MSE) and the error histogram as two performance metrics to assess the accuracy of various ANNs. The results reveal that optimal MSE outcomes were obtained when utilizing a balanced architecture with three hidden layers in the analysis of the sinusoidal-signal-based datasets.