Physics-Informed Neural Network-Based Adaptive Model Predictive Control for Visual Servoing of Robot Manipulators

In image-based visual servoing applications, the depth information of the feature points is inherently time-varying, which can lead to intractable computational challenges when conventional nonlinear model predictive control (NMPC) approaches are implemented. Furthermore, the depth information generally remains unavailable in vision-based measurements obtained through two-dimensional imaging systems. To address these challenges, a physics-informed neural network (PINN) is initially designed to compensate for these uncertain parameters minimizing the discrepancy between the linear system and the actual model. And that, by leveraging a more accurate model obtained through successive linearization, the adaptive model predictive control based on PINN (PINN-based AMPC) method is developed. The proposed control framework offers the advantage of generating relatively accurate system models with limited training data, while achieving rapid location of static object. Finally, the effectiveness of the proposed control method is demonstrated through a series of simulations conducted on a Universal Robots 3 (UR3) manipulator.