Real-Time Robotic Push-Grasping with DQN-Driven Transformer Attention Masking

Robotic arm grasping remains a fundamental yet challenging task in robotic manipulation, possessing significant research value. Efficient grasping in cluttered environments proves particularly difficult due to irregular object stacking and dense spatial distribution. We propose a push-assisted multi-action Deep Q-Learning Network (DQN) that enhances Q-network performance through Transformer's self-attention mechanism to capture global dependencies in both spatial and channel dimensions. The method improves training efficiency by utilizing adaptive threshold-generated mask maps and more accurately reflects object density through changes in the total length of Minimum Spanning Tree (MST) before and after pushing actions. Comparative experiments with the Visual Push Grasping (VPG) algorithm in the CoppeliaSim simulation environment, evaluated through our designed three-factor performance metrics and training process analysis, demonstrate that our approach not only enhances training efficiency but also significantly improves robotic grasping performance. Ablation studies further validate the effectiveness of our design.