An Inception Transformer-CNN Architecture for Robotic Grasping Detection

Addressing the issue of low accuracy in detecting suitable grasping regions for unknown objects through visual information, this paper proposes an Inception Transformer-CNN architecture for robot grasping detection, referred to as ITC-Grasp. The aim is to enhance the accuracy of grasping unknown objects. The architecture incorporates an Inception mixer with a channel-splitting mechanism, integrating the powerful capability of CNN in extracting high-frequency representations into the Transformer. By synergizing the local modeling capability of CNN and the global modeling capability of Transformer, it achieves superior feature representation, thereby improving grasping detection performance. Additionally, the architecture introduces a frequency slope structure to effectively balance high and low-frequency components across different layers. The results on the Cornell dataset show that the proposed method achieves 98.2% image segmentation accuracy and 97.1% object segmentation accuracy, respectively, showing competitive performance. The feasibility and effectiveness of the proposed method are verified by simulation experiments.