Learning processing skills from human demonstrations has become a growing research focus, particularly in robotic polishing. Current studies mainly emphasize trajectory reproduction, lacking task-level planning and the ability to generate precise position and orientation actions on freeform surfaces from visual input. To address this, we propose an action generation strategy tailored for robotic polishing on freeform surfaces. This method leverages RGB-D images to generate surface-aware actions. First, the 3D geometry of the workpiece is captured using an RGB-D camera and reconstructed into a structured triangular mesh. Then, a diffusion model infers continuous polishing actions from image observations. These actions are embedded onto the freeform surface using Mesh-DMP to ensure geometric consistency. The proposed framework integrates vision-driven policy generation with surface-constrained motion modeling, enhancing generalization and control in unstructured tasks. Experimental results demonstrate the effectiveness of our approach in generating high-quality polishing trajectories on complex surfaces.