PT-Grasp: coarse-to-fine affordance transfer for robotic grasp point prediction

This paper proposes a robot grasp point prediction method based on progressive transfer, named PT-Grasp(Progressive Transfer for Robotic Grasp Point Prediction), aiming to achieve high-precision grasping in complexenvironments. To address the problem of grasp point misalignment in existing methods caused by global pixel-wisealignment in cross-domain affordance transfer, we propose a two-stage, coarse-to-fine transfer approach. Firstly, the most relevant source reference is matched through hierarchical retrieval. Then, progressive affordance transfer isperformed on the target image. In the first stage of affordance transfer, the visual foundation model SSAMis usedtoestablish region correspondences between images, coarsely localizing the affordance region of the object (e.g., handles, edges). In the second stage of affordance transfer, the contact point is finely aligned within this region to generateanaccurate grasp point prediction. Extensive experiments on commonly used affordance grasp datasets showthat theproposed method significantly outperforms existing baselines, demonstrating stronger robustness and accuracy intaskssuch as open microwave, open oven, pick up bowl, and pick up bottle.