Research on Intelligent Grasping of Flexible Manipulator Combined with Visual Perception

This paper proposes an intelligent grasping system that integrates a deep learning-based visual perception method with a flexible manipulator control strategy. An improved YOLOv5 network and point cloud processing algorithm are used to achieve accurate detection and recognition of target objects. The system integrates RGB-D image data and object geometric information through an innovative adaptive feature fusion algorithm, and generates optimal grasping strategies via a deep reinforcement learning approach based on value iteration networks. This enables high-precision grasping of irregularly shaped, transparent, and reflective objects. Experiments conducted on 1,000 tests covering 50 different object categories demonstrate a grasping success rate of 93.857% under various lighting conditions and shape variations. Specifically, the success rates for irregular and transparent/reflective objects reached 90.125% and 87.652%, respectively, significantly outperforming existing methods. The system achieved a positional accuracy of 3.256 mm and a pose accuracy of 2.437°, representing improvements of 15.632% in grasping accuracy and 22.475% in execution efficiency compared to conventional approaches. Moreover, it exhibited excellent robustness under environmental disturbances, offering a practical technical solution for industrial automation production lines and service robotics applications.