Year: 2024
Type: article
Abstract: To address the issue of low grasp detection accuracy in robotic arms, an improved point cloud-based grasp detection model, TES-Net (TransformEdgeSAGE Network), is proposed, which builds upon the Edge Grasp Network architecture. In this model, PointTransformerConv layers are first employed to extract local features from the point cloud, integrating a self-attention mechanism to capture the complex relationships inherent within the point cloud data. Subsequently, the SAGEConv graph neural network module is utilized to perform feature aggregation on the adjacency graph, thereby mitigating the issue of neglecting the inter-point relationships during the feature extraction phase and enhancing the overall network performance. Experimental results demonstrate that, in comparison to existing state-of-the-art point cloud grasp detection methods, the proposed model exhibits superior generalization capability, improved robustness and stability, as well as higher accuracy.
DOI: 10.54097/9phdd715
Authors: Hongyi Li, Jianjun Zhu
Institution: Jilin University of Chemical Technology
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4405900852