Robotic Arm Grasping of Invisible Objects in Cluttered Environments

Grasping invisible targets using a robotic arm in cluttered environments remains a significant challenge.Many existing methods depend on redundant operations, which can compromise efficiency.In this paper, we introduce a novel framework called Deep Reinforcement Attention DenseNet (DRA-DenseNet), which utilises self-attention mechanisms for Q-value prediction.This framework excels at capturing global features of the objects, significantly enhancing the accuracy of action prediction in cluttered environments.We develop a search strategy that leverages the capabilities of DRA-DenseNet along with historical action data to locate invisible targets.Further, we propose an action coordination strategy to achieve more accurate grasping by training both the push network and the grasp network simultaneously.Experiments in both simulated and real-world environments demonstrate that the proposed approach outperforms many existing methods in terms of motion efficiency and success rate.