D2GNet: Efficient 6‐DoF Grasp Detection in Cluttered Scenes via Density‐Aware Dual‐Dimensional Graph Networks

ABSTRACT In real scenes, objects are often disordered, heavily stacked, and occluded, posing significant challenges to robotic grasping. To address these challenges, we propose Density‐Aware Dual‐Dimension Graph Network (D2GNet). Our model adopts a two‐stage architecture comprising a perception stage and an execution stage. In the perception stage, D2GNet employs a Dual‐Dimension Attention Module (DDAM) that integrates in‐degree‐enhanced graph attention (D‐GAT) and a Graph Channel Adaptive Attention (GCAA) to perceive local density and strengthen spatial reasoning in dense regions. In the execution stage, a target‐guided point‐cloud screening strategy extracts stable grasp candidates via multiscale structural enhancement and a rescoring mechanism. Experiments on GraspNet‐1Billion show that, even with limited training data, D2GNet consistently outperforms existing methods. Compared with the graph‐neural baseline GraNet, it achieves up to 16.83% improvement on RealSense evaluations and further exceeds the state‐of‐the‐art toward scale‐balanced 6‐DoF grasp detection by 2.16%, while delivering even larger gains on Kinect benchmarks. Real‐robot trials confirm its effectiveness; even under weak illumination, D2GNet exhibits strong robustness and clear potential for practical deployment.