Dynamics Modeling of Industrial Robotic Manipulators: A Machine Learning Approach Based on Synthetic Data

Abstract

Obtaining a dynamic model of the robotic manipulator is a complex task. With the growing application of machine learning (ML) approaches in modern robotics, a question arises of using ML for dynamic modeling. Still, due to the large amounts of data necessary for this approach, data collection may be time and resource-intensive. For this reason, this paper aims to research the possibility of synthetic dataset creation by using pre-existing dynamic models to test the possibilities of both applications of such synthetic datasets, as well as modeling the dynamics of an industrial manipulator using ML. Authors generate the dataset consisting of 20,000 data points and train seven separate multilayer perceptron (MLP) artificial neural networks (ANN)—one for each joint of the manipulator and one for the total torque—using randomized search (RS) for hyperparameter tuning. Additional MLP is trained for the total torsion of the entire manipulator using the same approach. Each model is evaluated using the coefficient of determination (R2) and mean absolute percentage error (MAPE), with 10-fold cross-validation applied. With these settings, all individual joint torque models achieved R2 scores higher than 0.9, with the models for first four joints achieving scores above 0.95. Furthermore, all models for all individual joints achieve MAPE lower than 2%. The model for the total torque of all joints of the robotic manipulator achieves weaker regression scores, with the R2 score of 0.89 and MAPE slightly higher than 2%. The results show that the torsion models of each individual joint, and of the entire manipulator, can be regressed using the described method, with satisfactory accuracy.