Abstract
Traditional control of robotic systems relies on the availability of an exact model, which assumes complete knowledge of the robot's parameters and all dynamic effects. However, this idealized scenario rarely holds in practice, as real-world interactions introduce unpredictable environmental influences, friction, and edge effects. This paper presents a novel data-driven approach to modeling and estimating robot dynamics by leveraging data collected during the robot's movements. The proposed method operates without prior knowledge of the system parameters, thereby addressing the limitations of conventional model-based control strategies in complex and uncertain environments. Our unified data-driven framework integrates classical control theory with modern machine learning techniques, including system identification, physics-informed neural networks (PINNs), and deep learning. We demonstrate its efficacy in the case of a two-link robotic manipulator that achieves superior trajectory tracking and robustness to unmodeled dynamics. The technique is modular and can be extended to manipulators with more joints.