Multi-Resolution Feature Fusion and Transfer Learning for 6D Grasp Pose Estimation of Category-level Object

Object grasping is fundamental for robotic interaction with the real world, yet remains challenging due to limitations in scene understanding and grasp generation. Traditional methods struggle with inaccurate RGB-based pose estimation, weak depth-based recognition, and the high cost of grasp annotation. This paper presents a novel 6D grasp pose estimation framework that fuses RGB and depth features at multiple resolutions via attention mechanisms, capturing rich correlations between color and geometry for improved category, shape, and pose estimation. A deformation-based model with transfer learning reconstructs point clouds of novel objects and transfers grasp configurations, enabling generalization. Extensive experiments show that the proposed method outperforms baselines and state-of-the-art approaches in both pose accuracy and grasp quality, while significantly reducing annotation requirements. This work offers strong theoretical contributions and practical potential for real-world grasp planning.