کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

SGG-Net: Skeleton and Graph-Based Neural Network Approaches for Grasping Objects

نویسندگان :

AliReza Beigy¹ Farbod Azimmohseni² Ali Sabzejou³ Mehdi Tale Masouleh⁴ Ahmad Kalhor⁵

1- دانشگاه تهران 2- دانشگاه تهران 3- دانشگاه تهران 4- دانشگاه تهران 5- دانشگاه تهران

کلمات کلیدی :

Robotic،Grasping،Grasp Pose Estimation،Graph Neural Networks،Straight Skeleton

چکیده :

Efficient and robust robotic grasping in cluttered, unstructured environments remains a critical challenge. Existing 6-DoF grasping techniques frequently rely on processing the entire observed point cloud, which can lead to high computational overhead and reduced precision. This paper introduces Skeleton and Graph-based Grasping Network (SGG-Net), an integrated framework that combines geometric skeletonization, implemented through the 3D-StSkel algorithm, with a Graph Neural Network (GNN) to effectively identify optimal grasp poses for robotic manipulation. The proposed method significantly narrows the search space by extracting and focusing on geometrically salient regions, enabling faster and more reliable 6-DoF grasp pose generation. Experimental evaluations demonstrate the approach’s strong performance across benchmark datasets, achieving grasp success rates of 90.74% on DexNet, 81.63% on EGAD, and 97.30% on YCB objects. Furthermore, the method outperformed state-of-the-art approaches on the GraspNet-1Billion dataset, achieving the highest Average Precision (AP) scores for both seen and novel objects. Experimental evaluations on multiple benchmark datasets, including DexNet, EGAD, YCB, and GraspNet-1Billion, show that the proposed technique achieves state-of-the-art grasp success rates and generalizes effectively to diverse and complex object shapes. This approach thus provides a scalable, accurate, and computationally efficient solution for grasp estimation, advancing robotic manipulation capabilities in real-world scenarios.