A Methodology of Stable 6-DoF Grasp Detection for Complex Shaped Object Using End-to-End Network

The proficient grasping of objects is a significant challenge in robotics, particularly in dynamic environments. Deep learning-based approaches have shown promise in adapting to changing situations and achieving successful grasping. Previous research utilizing deep learning to generate grasp candidates can be categorized into two approaches based on the degrees of freedom of a grasp: the 4-DoF (top-down) and 6-DoF methods. Due to the 4-DoF approach is limited by its lack of gripper orientation flexibility, the 6-DoF approach offers more accuracy and precision in grasping objects. This paper proposes improvements to the GSNet network, which is an open-source state-of-the-art network for 6-DoF grasping, through parameter tuning and the application of stable score and Multiscale Cylinder Grouping strategies. Detailed explanations are also provided on the method of applying different strategies to a single network and on the approaches of parameter tuning. By the implementation process, there were improvements for grasping complex-shaped objects and small objects. To validate the improvements, experiments were conducted to measure the AP in the scenes of the GraspNet-1Billion dataset. The results indicate that the maximum AP achieved in the case of novel objects is 27.92, which is higher than that of the original network. Additionally, the experimental results showed a success rate of 90.9% for bin picking with cluttered objects, demonstrating the practical utility of our network even in real-world environments.