A lightweight grasping pose estimation method for retail warehousing

Robotic grasping has been widely used in various industries. How to meet the requirements of grasping accuracy and grasping speed at the same time is a challenging problem in real-time grasping tasks. In this paper, aiming at the real-time grasping task in retail warehousing, a lightweight grasping pose estimation model for retail warehousing is proposed. The model first uses the Focus module to perform lossless double downsampling, and learns each feature map of the upper layer through the dilated convolution block to expand the receptive field; then, the R-Resblock structure is improved to perform multi-scale feature fusion, and a lightweight RFB-SE module is designed to enrich feature information and reduce the number of parameters. Finally, after upsampling and restoring the image, the grasping quality, grasping angle, and grasping width of the target are regressed to obtain the optimal grasping pose of the target item. Experiments are carried out in the Cornell dataset, Jacquard dataset, and simulation environment respectively. The experimental results show that the method has a grasping accuracy of 97.8% and a grasping speed of 78FPS on the Cornell dataset. The success rate is 91.5%, and the grasping task in a retail warehouse environment is simulated in grasping simulation experiments.