Adaptive Residual Physics-Informed Neural Network for Robust Dynamic Modeling

Physics-Informed Neural Network (PINN) has been applied to nonlinear robotic dynamics modeling. However, unstable parameter updates caused by measurement noise result in considerable degradation in model accuracy. Therefore, we propose an adaptive residual PINN (AR-PINN) to improve robustness in noisy conditions. First, we construct a residual connection kinematic computation layer to ensure the model parameters remain stable under noise fluctuations. Second, we incorporate dynamic equations into the feed-forward architecture to effectively guide the model parameter update. Experimental results on a two-axis robotic arm show that the proposed AR-PINN improves the accuracy of trajectory tracking with noisy data and remains stable even as the noise increases.