Research on Pixel-Level Grasp Configuration Prediction Method Based on Deep Neural Network

This paper proposes a pixel-level grasp configuration prediction method based on deep neural network. The method utilizes RGB images as inputs, combines with a deep neural network model, and outputs the object's grasp configuration at the pixel level. This paper adopts a new region-level AGA model to model the grasping properties of objects. The model solves the angle conflict during training and simplifies the process of marking the real grasping posture. A pose estimation network based on Deeplabv3 is designed to predict the OAR model on RGB images. Pixel-level mapping avoids the loss of real grasping posture and overcomes the limitations of current deep learning technologies by avoiding discrete sampling of grasp candidates and long computation times. Finally, experiments are conducted on the Cornell Grasp Dataset to verify the proposed method. The results show that the proposed method can accurately predict the grasp configuration of objects at the pixel level and has good predictive and generalization abilities.