AttEdge-PointNet++: Local Attention and Dynamic Edge Convolution for Robotic Peg-in-Hole Assembly

To address the requirements of accuracy and real-time performance in point cloud segmentation for robotic peg-in-hole assembly tasks, this paper proposes an improved AttEdge-PointNet++ network architecture. Building upon the PointNet++ segmentation framework, the model is enhanced by introducing a local multi-head attention mechanism and dynamic edge convolution. Specifically, a Transformer-based local multi-head attention module is embedded in the key sampling layer to strengthen the feature correlation of neighborhood point clouds. Additionally, a dynamic edge convolution operation is designed for the feature extraction stage, which enhances geometric detail representation through adaptive neighborhood weight learning. Experimental results demonstrate that the proposed model achieves a peak segmentation accuracy of 94.57% after 500 training epochs on the peg-in-hole assembly dataset, with both the mean Intersection-over-Union (mIoU) and instance-level IoU reaching 89.14%, showing significant improvements over the baseline model. While maintaining a lightweight structure, the network effectively improves the segmentation accuracy of target objects in assembly scenarios, providing reliable technical support for precise robotic assembly tasks.