Robot Dynamic Identification Based on Data Weighting Protection and Nonlinear Friction Separation

Dynamic identification is an important issue in robotics. Typically, using outlier data weighting and nonlinear friction models can effectively enhance the accuracy of dynamic identification. However, existing data weighting methods can cause data discontinuity and even regressor singularity. The modeling accuracy of the traditional Stribeck friction model can be further improved. Moreover, specifying arbitrary initial values for the nonlinear friction coefficients can decrease the overall identification efficiency of the convex optimization algorithm. To address these issues, this article proposes an identification framework based on data weighting protection and nonlinear friction separation. To deal with the outliers, a novel data weighting method is developed, which replaces outliers with the weighted average values of nearby nonoutlier data points, thus avoiding data discontinuity and regressor singularity. To improve the accuracy of the traditional Stribeck friction model, the arctangent function is used to replace the sign function, and the exponential velocity function is used to replace the linear one to describe the viscous friction phenomenon. In addition, a symmetric excitation trajectory is designed to separate the friction torques and then estimate the initial values of the nonlinear friction coefficients by data fitting, which can improve the computation efficiency of the proposed algorithm. Experimental results validate the correctness and superiority of the proposed algorithm. For instance, compared to other two methods, the proposed approach improves joint torque prediction accuracy by more than 12% and enhances identification efficiency by more than 92%.