Abstract: In order to solve the problem of low detection accuracy and long time when robot arms grasp unknown objects in dense scenes,a new grasping network structure called Cascade Fusion Grasp Detection (CFGD) was proposed,which was used to predict the grasping pose of objects in dense scenes. Based on the proposed backbone and several blocks for extracting initial high-resolution features,several cascade stages were also introduced to generate multi-scale features in CFGD network. A sub-backbone for feature extraction and an extremely lightweight transition block for feature integration were included at each stage.Such a design allowed for a larger proportion of parameters throughout the backbone and a deeper and more efficient feature fusion. Compared with the existing algorithms,the proposed algorithm has significantly improved the detection accuracy and speed on Cornell grasp data set,Jacquard grasp data set and custom data set. In the real grasping scene,the grasping success rate is 98.6 % in the single target scene and 94.6 % in the dense scene. Experimental results showed that the proposed algorithm can predict and grasp unknown objects in dense scenes with high accuracy.

Key words: mechanical arm, plane grasping, pose estimation, cascade fusion, machine vision