No abstract

This paper considers comparative evaluation of recent methods for grip point determination for manipulations with objects in the scene. This research is aimed to compare and evaluate the modern approaches of grip point determination, when this process is aided by computer vision. The methods of object gripping, considered in this paper, are employed in connection with depth map composition, backed by neural network model ResNet-50, which allowed to omit application of specific depth sensors in the course of experiments. This research shows dependencies of successful grip probability from the things being manipulated. Probability scores, averaged over different types of objects for the methods GPD, 6-DOF GraspNet, VPG, were, accordingly, 0.690, 0.741, 0.613. The paper also considers dependencies of successful grip probability from object sizes, distances from the capturing camera and target objects in the scene, luminosity levels, as well from the angles of scene inspection along the vertical axis. In terms of the considered methods GPD, 6-DOF GraspNet, VPG, non-linear increasing dependencies are revealed for object type-averaged probabilities of successful grip from luminosity level of the scene. It was also discovered, that the dependencies of successful grip for all the other parameters are non-linear and non-monotonic. The ranges of the values for scene parameters under consideration are defined in this paper, which ensure the highest probability values for object grip in these approaches. Upon the results of the performed experimental evaluation, the 6-DOF GraspNet solution showed the best performance for the vast majority of the considered parameters of the scene. The approach, presented in this paper, is the preferable way for solution of grip point problem, in context of methods, which assume depth map reconstruction without specific equipment.

In development of formation control systems for modular robotic devices, especially relevant is the problem of analysis and, particularly, segmentation of complex surfaces, down which the robotic device will move. An approach based on fine-tuning of a neural network model HRNet was developed to solve the problem of segmentation of complex surfaces. Model fine-tuning was performed based on a custom dataset, which included 15,000 labeled images. The training dataset included the scenes of the following types: scenes with stairways, scenes with even surfaces, scenes with isolated obstacles and with groups of obstacles. Approbation and functional quality assessment of the developed approach were performed based on the test dataset, which included 3000 images with different levels of scene illumination. According to the results of the testing, the developed approach shows decent quality of segmentation on images with even surfaces (IoU = 90.2%) and with stairways (IoU = 71.3%) as well maintains some resilience to the variations in the scene luminosity levels. After fine-tuning, the averaged performance metrics of this neural network model on images with luminosity levels of 100% and 70% increased by 9.1% and 7.3% at average and resulted in 65.5% and 52%, respectively.

No abstract

No abstract