A Novel Robot Collision Detection Method by Predicting and Compensating External Torque Residual Considering Frequency Distribution

The generalized momentum observer (GMO) is widely used for robot collision detection, but its accuracy is significantly affected by external torque residuals caused by dynamic model errors. However, existing studies lack residual compensation strategies specifically designed for the GMO framework. In particular, friction modeling errors in the velocity reversal region severely degrade detection accuracy, and their strong coupling with other residual components (e.g., gravity, Coriolis, centrifugal forces) further increases prediction difficulty. This paper proposes a machine-learning-based external torque residual prediction and compensation method tailored for GMO. First, representative input features are derived through theoretical analysis of the residual formation process, ensuring clearer mapping to residuals and improving prediction accuracy. Second, a frequency-distribution-based residual decomposition method (RD-FD) is introduced to isolate friction-related errors in the velocity reversal region, enabling targeted prediction and enhancing overall estimation accuracy. Finally, an XGBoost-based residual prediction model is developed, and a compensation framework is constructed to refine external torque estimation. Additionally, a time-varying threshold design is proposed to mitigate accuracy loss caused by friction modeling errors, further enhancing collision detection performance. Experimental results show that the proposed method accurately predicts and compensates GMO residuals, significantly enabling lower detection thresholds, which enhances sensitivity and reduces detection delay.