A context-aware grasp detection framework for robotic grasping

Abstract

Robotic visual grasping in cluttered scenes is more challenging than grasping a single object in structured environments due to the complex inter-object interactions, which lead to visual occlusion, object stacking and detection ambiguity between similar objects. Existing approaches deal with the problem by designing advanced representation networks for an accurate prediction of gripper configurations, disregarding the physical contacts of objects. To this end, a context-aware grasp detection framework considering the spatial relationships of objects is proposed. Specifically, our framework represents cluttered scenes using a graph convolutional network to depict the global scenario context and topology of objects. By fusing pyramid attention modules into the network, local core features are extracted and strengthened, guaranteeing the effective sequential grasping of stacked objects. Experiments on the Jacquard dataset and real-world scenes demonstrate our method.