CGNet: Robotic Grasp Detection in Heavily Cluttered Scenes

Robotic grasp technology has been widely used. However, the robotic grasp in cluttered scene is still a challenging problem. In this regard, this article proposes a novel robotic grasp detection method cluttered grasp network (CGNet). First, to make the network fully focus on important features, this article proposes a novel attention module two branches squeeze-and-excitation residual network (TSE-ResNet) and uses it as the backbone to extract features. Then, to detect grasp rectangle more accurately, a novel grasp region proposal module is proposed, which can well utilize the multiscale features and refine the grasp region. Finally, a novel position focal loss is proposed to detect the rotation angle of the grasp rectangle, and can well solve the problem of discontinuous rotation angle. The CGNet is trained and tested on the GraspNet-1Billion dataset and Cornell dataset, achieving 87.9 and 97.9% accuracy, respectively. Moreover, to test the effectiveness, the CGNet is also tested on the Multiobject dataset and Clutter dataset. The detection results show that the CGNet can well detect the grasp rectangle when faces unseen objects. The ablation experiments are also performed to verify the performance of proposed modules. The experimental results show that the proposed modules can improve the detection accuracy in the cluttered scene. Finally, to evaluate the generalization of the CGNet, it is also evaluated in the real world, and applied to the grasp task of a real Baxter robot, and obtained a grasping success rate of 91.7%.