<p>Robotic grinding provides a flexible approach for machining complex freeform surfaces, but accurate material-removal prediction and reliable execution compensation remain challenging because of contact variation, robot response delay, and force–feedrate inconsistency. To address these issues, this study proposes a Residual-Corrected Incremental Learning Framework (RILF) and deploys the learned predictor to force–feedrate coordinated compensation for robotic grinding. First, a Hertzian-based mechanistic baseline is established to describe the dominant material-removal trend. The physical sources of prediction residuals, including contact-point shift, contact-patch distortion, stress redistribution, and curvature-dependent parameter coupling, are then analyzed. Based on this analysis, a GWO-XGBoost residual learner is constructed to compensate for the systematic deviations of the mechanistic baseline. A staged incremental updating mechanism is further introduced to incorporate newly available process data while maintaining stable prediction performance. Second, the learned predictor is deployed to an execution-oriented compensation layer, where feedrate-deviation-driven force correction is used to reduce material-removal errors caused by robot motion lag and feedrate mismatch. Experimental results show that the proposed framework improves material-removal prediction accuracy, enhances compensation performance, and maintains acceptable feedrate and force tracking accuracy during actual robotic grinding. The proposed method provides a coherent prediction-to-execution framework for improving profile accuracy and process stability in robotic grinding of complex freeform surfaces.</p>