Enhancing Robotic Grasping Pose Estimation with Channel Interaction Attention Refinement

Robotic grasping remains a significant challenge within the field of robotics, particularly in complex, dynamic, and highly varied environments. To improve the accuracy and robustness of grasping tasks, we propose a novel 7-degree-offreedom (DoF) grasping pose estimation framework with an encoder-decoder architecture and sparse convolutional network. This framework effectively captures both global and local contextual information from the input point clouds, providing a solid foundation for subsequent grasping decisions. Additionally, we introduce the channel interaction attention refinement mechanism to enhance multi-scale non-similarity relationships among feature channels. Experimental results on the GraspNet-1Billion dataset show that our method significantly outperforms existing state-of-the-art approaches, demonstrating strong accuracy for robotic grasping tasks.