Optimization of Dual-Constraint Robotic Arm Grasping Operations Based on an Optimized Contact Grasping Network

Abstract

With the continuous advancement of intelligent manufacturing and robot perception capabilities, traditional robotic arm grasping methods still face problems such as inaccurate pose prediction and poor task adaptability when operating in dynamic scenes, handling complex objects, and executing functional action. Therefore, this research develops an optimized contact grasping network model that integrates scene and task constraints. It combines a UR5 six-degree-of-freedom robotic arm, visual input, and the Contact-GraspNet architecture based on point cloud, and introduces a PointNet++ local feature enhancement mechanism and
a lightweight encoder design to effectively improve the spatial perception and action planning capabilities of grasping points. Experimental results show that, on the GraspNet and YCB datasets, the model achieves F1-scores of 92.54% and 91.82 %, respectively, with the average execution time reduced to 0.61 s. In functional operation scenarios involving door handles, kettle handles, and drawer pulls, the grasping accuracy remained above 0.89, with the task completion rates significantly outperforming those of mainstream baseline models. Under visual interference conditions with an occlusion rate of up to 75 %, the average inference latency
was controlled within 0.82 s. Under varying light intensities, the pose angle error remained within the range of 1.21° to 1.87°. Therefore, this model exhibits comprehensive advantages in grasping accuracy, latency control, and deployment efficiency, and has the potential to be largely applied in industrial, service, and specialized task environments.