Improved Cascade RCNN Model-Based Research on Multi-Object Grasping Technique for Robotic Arms

Robot multi-object grasping in complex environments is a hot topic in current technology research. While deep learning-based single-object grab detection has progressed, multi-object scenarios remain challenging. This paper proposes a Cascade R-CNN network model based on RSEN module optimization to im-prove the efficiency and accuracy of multi-target capture detection. Firstly, we extended the Cornell Dataset and constructed a multi-objective capture dataset (MOGD) to enhance the model's generalization ability. Then, the RSEN module is introduced to design the RSEN-Cascade R-CNN network. The five-parameter method describes the two-dimensional grab frame and verifies it on the new data set. Experiments show that the model shows performance convergence after 60 epochs, the capture detection time is only 14 ms, and the accuracy is as high as 94.8%. In the experi-ment of grasping objects on the Kinova robot arm, the method achieved a high grasping success rate of 91.6%. This research is innovative in theory and shows significant application potential in automated production lines, logistics sorting, and service robots. By optimizing Cascade R-CNN and introducing the RSEN module, the robot's grasping and detection performance in a multi-target environment is significantly improved.