关键词: 抓取检测, 点云Transformer, 多尺度特征, 尺度感知

Abstract: Grasping detection is fundamental to achieving precise robotic grasping. However, existing grasping detection networks usually cannot adequately extract point cloud features, meanwhile multi-scale information of sampled points is usually ignored, furthermore imbalanced sample distribution usually cannot avoid in most datasets. To address these issues, a multiscale grasping detection network model integrating point cloud Transformer is proposed. An encoder-decoder network integrating point cloud Transformer is constructed, which utilizes self-attention mechanism to enhance information interaction between local feature vectors. A multi-cylinder scale perception module is designed to capture the local features of different receptive fields of the sampled points, so that the proposed model can adaptively select effective features to capture the multi-scale information of sampled points. A scale-balanced learning loss function is designed to alleviate the scale bias phenomenon during the training process through reasonable weight assignment. The proposed model is validated on the GraspNet-1Billion dataset, demonstrating that its performance surpasses existing models. Grasping experiments in real-world scenarios further confirm the strong generalization ability of the model, contributing to stable and efficient precise grasping in practical applications.

Key words: grasping detection, point cloud Transformer, multiscale features, scale perception