Abstract Accurate modeling of robotic manipulator dynamics is vital for achieving precise and robust control, particularly under nonlinear and coupled motion conditions. This study explores the capability of machine learning techniques to capture the nonlinear dynamics of an articulated manipulator and predict joint torques with high fidelity. Four function approximation schemes—Artificial Neural Network—Bayesian Optimization (ANBO), Support Vector Machines (SVM), Gaussian Processes (GP), and Decision Trees—are optimized using Bayesian hyperparameter tuning and systematically evaluated. The results show that: (1) among ANBO architectures, the multiple-input single-output (MISO) configuration achieved the highest prediction accuracy; (2) ANBO consistently provided reliable torque estimation across all joints, while GP offered comparable performance except at low torque magnitudes; and (3) Decision Trees and SVMs yielded lower accuracy, reflecting limited ability to capture complex nonlinear behaviors. Overall, the findings demonstrate the effectiveness of ANBO-based models for adaptive inverse dynamics estimation and highlight their potential for future exploration of adaptive dynamic modeling, particularly in scenarios where variable payloads are involved to further optimize the performance and control of many robotic systems.
Yusuf et al. (2026) studied this question.