A Residual-Driven Decomposed PINNs Method for Dynamics Identification of Robot Manipulators

In dynamic parameter identification for robots, traditional least squares (LS) methods exhibit limitations in capturing complex nonlinear friction characteristics. To address this challenge, we propose a residual-driven decomposed Physics-Informed Neural Networks (PINNs) framework that synergistically integrates physics-based identification with data-driven learning. The methodology follows a two-stage approach: First, conventional LS estimation provides initial dynamic parameter values. Subsequently, the PINNs architecture employs a hybrid loss function combining physical loss and data loss to refine identification accuracy through residual minimization. Experimental validation demonstrates that our approach achieves a reduction in torque prediction root mean square error (RMSE) compared to the baseline method, confirming its superior performance in robot dynamics identification.