This paper presents an efficient cascade calibration method with an improved Levenberg–Marquardt and sine–cosine hybrid algorithm to enhance the absolute positioning accuracy of robotic grinding systems. To expedite convergence in the Levenberg–Marquardt algorithm, a dynamic adaptive weight mechanism is introduced, enhancing global and local search capabilities. Furthermore, a novel learning rate, combining exponential and cosine functions, addresses local optima in the algorithm. The improved Levenberg–Marquardt algorithm is employed to obtain suboptimal values for robot kinematic parameter deviations. Subsequently, these values are used as central points for generating a candidate solution set in the sine–cosine algorithm, resulting in more accurate kinematic parameter deviation identification. This innovative dual-search optimization approach combines the two algorithms. Experimental results confirm the substantial improvements in absolute positioning accuracy and surface machining precision achieved by the proposed model, with the calibration method’s effectiveness verified through experimentation.