Industrial robots, leveraging their high flexibility and reconfigurability, are widely applied in grinding large complex components. But for variable defects like splattered welding slags on automotive body-in-white, existing robotic grinding strategies cannot adjust dynamically, causing unstable quality. To effectively remove random defects, this paper proposes an intelligent process decision-making method via incremental learning and database. First, orthogonal experiments and F-test are employed to screen out effective modeling data. Subsequently, an incremental support vector regression (SVR) algorithm is developed to establish an upgradable robotic grinding roughness model and material removal model. On this basis, a weighted case-based reasoning (CBR) method combined with an improved multi-objective grey wolf optimizer (MOGWO) is employed to determine optimal process parameters under complex grinding conditions. Meanwhile, the residual height of weld slags after grinding is predicted by the material removal model, thereby supporting secondary grinding and reducing the need for repeated measurements. Experimental results demonstrate that random weld slags can be effectively removed. The average surface roughness (Ra) of the processed areas can reach 0.95 μm, with an average prediction error of only 9.0%, validating the effectiveness of the proposed intelligent process decision-making method.