IMPROVING GRASP POSE GENERATION USING GGCNN AND PRINCIPAL COMPONENT ANALYSIS

Andreev Andrey Valeryevich

The article is dedicated to the development of a modified grasp pose generation algorithm based on GGCNN with the integration of principal component analysis (PCA) to improve the accuracy of robotic object grasping without retraining the neural network. This problem is highly relevant given the evolving methods of machine vision and neural network approaches in robotics. Working with three-dimensional images is increasingly popular as a replacement for classical machine vision techniques and two-dimensional image processing. As robotic systems and mobile robots advance, a challenge arises in grasping objects that are often randomly located within the robot’s operational workspace. This necessitates the use of advanced methods for object detection and subsequent grasping. Equally important is the application of specific algorithms for seemingly simple tasks such as picking up, transferring, and sorting objects in industrial production. This work proposes an enhanced approach that combines the generative grasping convolutional neural network (GGCNN) with PCA analysis. The methodology includes preprocessing of input data—namely the depth map from an Intel RealSense camera—analyzing the object grasp probability map using PCA, and transforming the data into a format suitable for grasp execution. Experimental data were collected using a pretrained GGCNN on the Cornell dataset to obtain metrics for grasp success. As a result of this work, an acceptable grasp success probability of approximately 95.6% was achieved for the test object by applying PCA to the parametric grasp probability map. The study also evaluated grasp probabilities under various point-selection parameters for PCA when dealing with a symmetrical object. The proposed method increases the object grasp success probability without retraining the neural network, thereby reducing the labor required to create and label training datasets and retrain the network for new objects. It also reduces the number of output parametric maps and, consequently, decreases the dimensionality of the network’s output.

Keywords:principal component analysis, PCA, robotic grasping, GGCNN, convolutional neural networks, CNN, manipulation tasks, computer vision, robotics.