A graph reasoning method for multi-object unordered stacking scenarios

• Published in: 2023 IEEE International Conference on Real-time Computing and Robotics (RCAR) pp. 888 - 892
• Main Authors: Zuo, Guoyu, Wang, Zihao, Gong, Daoxiong, Huang, Gao
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.07.2023
• Subjects: BiFPN; Cognition; Computational modeling; EfficientNet-B0; Feature extraction; graph attention networks; Grasping; intelligent robot grasping; Real-time systems; Stacking; Visualization
• DOI: 10.1109/RCAR58764.2023.10249310

In order to enable the robot to safely pick the target object from it in the face of a scene with multiple objects stacked in a disorderly manner. It is of great importance for intelligent robot grasping to get the operation order by obtaining the position relationship between the target object and other objects in space. In this paper, an end-to-end detection inference network is proposed, in which features are extracted from the input RGB information by EfficientNet-B0 combined with BiFPN in terms of feature extraction, and in terms of operation relationships, object pairs without up-down position relationships are first eliminated to reduce the amount of operations, and graph attention networks (GAT) are later used to reason about the positions between objects. The network model is trained and tested on the VMRD dataset, and the experiments show that this model has good results in inference of object spatial position relationships for multi-object unordered stacked scenes.