Robotic Grasp Pose Estimation Oriented Deep Learning Algorithm Based on Residual Network

Autonomous manipulator grasp is an important issue in robotics research. To obtain the optimal grasp pose, we combine manipulator vision and deep learning to realize the artificial intelligence of the manipulator grasp. We adopt the idea of using residual network to improve the generative grasping convolutional neural network (GG-CNN). Firstly, we build a convolution residual module. By piling multi-layer of residual modules, we can build the residual network and deepen the depth of the convolutional neural network, which is used as the main part to improve GG-CNN. Improved GG-CNN based on deep residual network enhances the accuracy of the optimal grasping pose generation of the manipulator. Experimental results show that the accuracy of the improved GG-CNN model based on residual network reaches 88%, which is much higher than the original model's accuracy of 72%. It significantly improves the accuracy of the model to predict the optimal grasp pose of the manipulator.