Robotic Arm Target Grasping Algorithm Based on Obstacle Removal in Space-Constrained Scenes

Achieving precise grasping of specific targets in complex and crowded scenes is of great importance. Existing works often rely on the combination of pushing and grasping strategies to separate the target from surrounding obstacles. However, in crowded environments with limited space, traditional combined push-grasp methods struggle to effectively separate the target from dense obstacles due to insufficient action execution space. To address the problem of target grasping in space-constrained scenes, this paper proposes a target grasping algorithm based on policy search mechanism. By intelligently identifying and grasping the obstacles around the target, the complexity of the workspace is reduced, thereby improving the action efficiency and the success rate of target grasping. Firstly, a Q-value prediction network is trained based on deep reinforcement learning to extract features from input images and predict the Q-value of each pixel point to determine the best grasping point. Secondly, to improve the network's prediction ability, the SE-DenseNet network structure is proposed for feature extraction and enhancing the network's sensitivity to key feature information. Finally, to verify the effectiveness of the proposed grasping framework, a space-constrained working scene is built in the CoppeliaSim simulation platform for experimental testing. The results show that the proposed framework achieves an action efficiency of 82.6% and a grasping success rate of 83.5%, significantly outperforming traditional methods. Moreover, the impact of each part of the algorithm on system performance is further verified through ablation studies, fully demonstrating the efficiency and reliability of the proposed framework.