Fauzy Satrio Wibowo, Hsien-I Lin, Wen-Hui Chen, Siti Sendari
This paper presents a comparative study of three inverse dynamic identification methods: ordinary least squares (OLS), weighted least squares (WLS), and recursive least squares (RLS). The experiment was applied to a 2-DOF robotic manipulator under two excitation conditions. Identification Trajectory 1 ('ID1') comprises ten harmonic sinusoids and, while Trajectory 2 ('ID2') spans a wider frequency band. We quantify the fidelity of the static parameter estimation via the mean absolute error in nine basic inertial and friction parameters, and assess the precision of the dynamic torque prediction using RMSE and MAE in both the identification and the held test trajectories. Under limited excitation, WLS achieves the lowest parameter error and joint RMSE (0.1173 Nm), closely followed by OLS. Under rich excitation, OLS produces the most accurate parameter estimates, and RLS attains the lowest joint RMSE. These results demonstrate that increasing trajectory richness improves both static parameter recovery and dynamic torque prediction performance. © 2025 IEEE.
Universitas Negeri Malang, Department of Electrical and Informatic Engineering, Malang, Indonesia; National Yang Ming Chiao Tung University, Institute of Electrical and Control Engineering, Hsinchu, Taiwan; National Taipei University of Technology, Graduate Institute of Automation Technology, Taipei, Taiwan
