FGNet: Faster Robotic Grasp Detection Network

Abstract

In unstructured environments, achieving fast and accurate object detection and successful grasping presents a significant challenge. Current grasping detection networks primarily focus on reducing the network's floating point operations (FLOPs) to improve network speed. However, we found that the effectiveness of this method is not particularly significant. To address this issue, we employed partial convolution (PConv) in place of regular convolutions to significantly enhance the network's detection speed. Additionally, we implemented a parallel structure for the network to fuse low-level and high-level features, reducing the loss of detail information during the decoding process. Our proposed faster grasp detection network (FGNet) achieved a performance of 96.74% (ow) and 98.66% (iw) on the Cornell dataset, with a detection speed of only 11ms. The grasping success rate was 96.5% in single-object scenarios and 92% in cluttered grasping scenarios.