FCNN-GraspNet: a steamlined neural network for robotic grasp detection

With the development of robotics, generative grasp detection network based on deep learning has become a hot research direction in the field of robotic arm grasping. The existing learning-based grasp detection methods usually cannot guarantee high accuracy while keeping their network models small. To address this issue, in this paper, a lightweight grasp detection neural network FCNN-GraspNet is proposed. It is designed using the encoder-decoder structure in a neural network By inputting an n-channel multi-modal image, a generate pixel-level grasping detection frames. In the encoding layer, partial convolution is used to extract the features of the image, and in the decoding layer, CSPBlock and upsampling network are designed to predict the grasping pose for the grasping generation head network. Its network training parameter is only 0.947million. The experimental results show that this network achieves up to 99.3% accuracy on the Cornell dataset and 92.7% accuracy on the Jacquard dataset. Meanwhile, Achieved an accuracy of 83.7 in 200 grasping experiments in a simulated environment.