Key points are not available for this paper at this time.
Accurate grasping of everyday objects is crucial to the development of intelligent home service robots. However, the existing mechanical gripper grasping methods have a low grasping success rate and poor adaptability for everyday objects, making it difficult to meet the requirements of intelligent home service robots. To address this issue, different from using conventional visual information or other sensor information to identify objects, this paper proposes a novel soft gripper grasping method based on 3D point cloud upsampling to effectively grasp everyday objects. First, the PU-GCN (Point Cloud Upsampling using Graph Convolutional Networks) is improved by integrating the modified multi-scale parallel feature extraction module and a self-attention module. It makes the shape and contour of the object more realistic to improve the object’s grasping accuracy. Then, by combining the upsampled 3D point cloud information, a cross grasping strategy based on object aspect ratio constraints is designed. It significantly enhances the adaptive capability of the soft gripper. Experimental results show that the proposed grasping method outperforms other grasping techniques in terms of adaptability, enabling accurate grasping of various everyday objects. The improved PU-GCN up-sampling method has a better grasping ability than the original PU-GCN, and its grasping success rate is improved from 92.7% to 95.3%. The grasping success rate and grasping completion rate of the proposed cross grasping strategy are 10.7% and 4% higher than that of the random grasping strategy, respectively. This novel approach provides intelligent service robots with a promising solution for effectively grasping everyday objects.
Zhang et al. (Tue,) studied this question.