Enhancing Pixel-Wise Robotic Grasping with Residual Blocks

This paper presents an effective neural network, Residual Generative Grasping Convolutional Neural Network (ReGG-CNN), to predict reliable robotic grasps for novel objects based on an n-channel scene image. Based on Generative Grasping Convolutional Neural Network (GG-CNN) architecture, ReGG-CNN incorporates residual blocks to enhance performance while preserving the network’s lightweight and single-pass generative nature. ReGG-CNN is evaluated on two standard open-source datasets, Cornell and Jacquard grasping datasets, achieving 87% and 89% accuracy, respectively. Moreover, multiple cluttered scenes are created using novel household and adversarial objects to assess the model’s capability to generalize to most kinds of objects and make multi-grasp predictions. Even though only training on individual objects, ReGG-CNN effectively predicts grasps for diverse objects in cluttered scenes.