Antipodal-points-aware dual-decoding network for robotic visual grasp detection oriented to multi-object clutter scenes

Tian, Hongkun; Song, Kechen; Xu, Jing; Ma, Shuai; Yan, Yunhui

doi:10.1016/j.eswa.2023.120545

Cited by 1 publication

(1 citation statement)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Ni et al [21] proposed an end-to-end bit position estimation network based on PointNet++ using sparse point cloud as input and Ferrari-Canny multi-target pickup algorithm, the result of robot grasping accuracy is higher than PointNet. However, the deep learning method involves the labeling of datasets and a large amount of computation in the training process, the robustness and stability are poor, and it is not suitable for industrial production [22,23,24,25].…”

Section: Introductionmentioning

confidence: 99%

A 3D pickup estimation method for industrial parallel robots based on point cloud simplification and registration

Zhang,

Wang,

Wang

et al. 2023

Preprint

View full text Add to dashboard Cite

In industrial environments, cameras are strongly affected by light and viewpoints, and parallel robots based on traditional vision methods have poor sorting abilities. In two-dimensional vision, depth information cannot be estimated, and parallel robots can only pick up objects based on the fixed height. In this paper, we propose a 3D pickup estimation method for parallel robots based on point cloud simplification and registration for multi-objective pickup tasks. Firstly, a point cloud segmentation method based on the spatial domain is proposed to separate incomplete object point clouds from robot space. The 3D scanner scans the complete object point clouds to produce the object dataset. Secondly, according to the fast point feature histogram (FPFH) and the weight locally optimal projection (WLOP), a fusing the FPFH and WLOP (FF-WLOP) method is proposed to simplify the incomplete point cloud and obtain more distinctive edge features of objects. The complete point cloud in the dataset is aligned with the simplified incomplete point cloud and the coordinate information of the calculated barycenter is given to the incomplete point cloud. Then a dynamic weight singular value decomposition (D-SVD) hand-eye calibration method and a best projection point strategy are proposed to transform the barycenter coordinates of the object to the best pickup coordinates. The experiment results show that the point cloud registration error is 0.38mm, the pickup rate is 92%, and the robot positioning error is 4.67mm, which meets the basic pickup requirements.

show abstract