Investigating Neural Network Hyperparameter Variations in Robotic Arm Inverse Kinematics for Different Arm Lengths

This research delves into the evaluation of Artificial Neural Networks (ANNs) in tackling the intricate domain of inverse kinematics problems within the context of robotic systems. Compared to direct kinematics, inverse kinematics presents significant challenges, especially when dealing with different arm lengths, which can impact the system's overall functionality. In this study, we conduct a comprehensive comparative analysis to assess the performance of ANNs across various robotic arm lengths. Following the forward kinematics of a simple 2-DoF robotic manipulator, three distinct datasets (fixed step size, random step size, and sinusoidal step size) are generated for training, testing, and validation phases. For every dataset, three optimizers, namely Levenberg Marquardt, Bayesian Regularization, and Scaled Conjugate Gradient, are considered with different hyperparameters to investigate the performance of the ANN models. Based on the validation of MSE values, ANN models with different hyperparameters are found to be a promising approach in estimating the inverse kinematic solutions for different arm lengths with data complexities. Such solutions can be further extended to higher DoF robotic solutions and save the computational burden of evaluating analytical solutions.