Feedforward Backpropagation Artificial Neural Network for Modeling the Forward Kinematics of a Robotic Manipulator

Determining an end-effector position and orientation, known as forward kinematics (FK), is a crucial aspect of robotic manipulation. This research addresses the FK problem using a Feedforward Backpropagation Artificial Neural Network (FFBP-ANN) and examines the network’s performance with different hyperparameters. The input of the FFBP-ANN is primarily formed using three different datasets: fixed-step values, random values, and sinusoidal-signal based values, for a 2-DOF robotic manipulator. The output of the FFBP-ANN is evaluated using FK formulation of the robot manipulator. Thereafter, the FFBP-ANN is trained utilizing the input-output training dataset using The Neural Network Toolbox for MATLAB. The network’s potential is investigated by varying the critical hyperparameters, including the number of hidden neurons and the number of hidden layers. Moreover, the effectiveness of the proposed FFBP-ANN is investigated with three different training optimizers including the utilization of Levenberg-Marquardt, Bayesian Regularization, and Scaled Conjugate Gradient optimization algorithms. During training, by adjusting the network’s weights and biases iteratively, the backpropagation technique allows the network to effectively learn and comprehend the relationship between joint positions and Cartesian coordinates. The comparative results offer valuable insights into the strengths and limitations of different ANN architectures when applied to solving FK problems. In this way, this study focuses on exploring the application of the FFBP-ANN in resolving FK challenges and assisting in the identification of appropriate network types for path-planning tasks.