A PINN-Based Friction-Inclusive Dynamics Modeling Method for Industrial Robots

Abstract

High-precision dynamics and friction models are crucial for high-performance control and operation of industrial robots. However, due to the requirement for model linearization, mainstream identification-based modeling methods struggle to capture nonlinear features of the model. In recent years, physics-informed neural network (PINN)-based methods have achieved interpretable nonlinear robotic dynamics and friction modeling, but suffer from suboptimal accuracy due to the lack of comprehensive modeling and learning strategies. This article presents a PINN-based friction-inclusive dynamics modeling method for industrial robots. A hybrid learning strategy for robot dynamics and friction is designed, ensuring modeling accuracy while avoiding reliance on joint torque component labels. Furthermore, residual error compensation is integrated into the proposed PINN to enhance its capability to learn nonlinear features. Experimental validation on two different robots demonstrates the effectiveness of the proposed method. Compared with other advanced methods, the average joint torque error is reduced by an average of 39.69%.