Physics-Informed Neural Network-Based Input Shaping for Vibration Suppression of Flexible Single-Link Robots

Abstract

The vibration suppression of flexible robotic arms is challenging due to their nonlinear spatiotemporal dynamics. This paper presents a novel physics-informed neural network (PINN)-based input-shaping method for the vibration suppression problem. Through a two-phase training process of a neural network based on a loss function that follows both the physical model constraints and the vibration modal conditions, we identify optimal input-shaping parameters to minimize residual vibration. With the use of powerful computational resources to handle multimode information about the vibration, the PINN-based approach outperforms traditional input-shaping methods in terms of computational efficiency and performance. Extensive simulations are carried out to validate the effectiveness of the method and highlight its potential for complex control tasks in flexible robotic systems.