Physics-informed neural networks for compliant robotic manipulators dynamic modeling

Abstract

Deep learning is widely used in robotics, yet often overlooks key physical principles in dynamic modeling, leading to a lack of interpretability and generalization. To address this issue, recent innovations have introduced physics-informed neural networks (PINNs), which integrate fundamental physics into deep learning and offer significant advantages in modeling rigid-body dynamics. This study focuses on the application of PINNs to model compliant robotic manipulators. This requires extending PINNs to handle complex compliant dynamics. We propose an augmented PINN model capable of comprehensively learning manipulator dynamics, including compliant components. The model is tested on dynamic modeling of two physical compliant manipulators and a simulated manipulator. The results highlight its exceptional precision and generalization across a wide range of robotic systems, from purely rigid to compliant structures.