Improved GR-Convnet for Antipodal Robotic Grasping

This paper introduces a robot system designed to address the problem of performing antipodal robot grasping for unknown objects. We focus on the high-level approach of GR-Convnet for the task and propose a neural network with high robustness while maintaining real-time performance. The three improvements include introducing Squeeze and Excitation (SE) blocks, removing Dropout in the final layer, and using Residual Block and Concurrent Spatial and Channel Squeeze and Channel Excitation (scSE) Block. We evaluate the proposed network on the Jacquard dataset containing information on various household objects. As a result, we achieved an approximately 7.2% improvement in accuracy compared to GR-Convnet. Additionally, using a real robot, we demonstrated a grasp success rate of 93.3% and 92.5% for household and adversarial objects, respectively.