Abstract
To address the challenge of grasping multi-target objects with uncertain shape, attitude, scale, and stacking, this study proposes a high-performance planar pixel-level grasping network called random sub-path grasp fusion network (RSPFG-Net). The paper introduces the agile grasping representation (AGR) strategy for dexterous grasping of target objects and constructs a Multi-objects Grasping Dataset (NEU-MGD). Secondly, the article introduces the Multi-Scale random sub-path fusion (MSRSPF) module. This module effectively prevents overfitting and improves the robustness of the grasping network in unstructured scenes. The MSRSPF module is connected with the DeepLab v3 network to form the RSPFG-Net for pixel-level grasping and multi-target high-performance grasp detection. Finally, the experiments conducted with RSPFG-Net on publicly available Cornell, Jacquard, and NEU-MGD datasets resulted in an average grasping detection accuracy of 97.85%. In real-world scenarios, the robot achieved an average grasping success rate of 94.31%. These results demonstrate the excellent performance and robustness of RSPFG-Net when it comes to multi-target grasping problems.