Learning based robotic grasping of objects

Abstract

In recent years, the rapid development of artificial intelligence and machine learning has brought significant advancements and opportunities across various industries. The integration of robotics and deep learning has enabled robots to perform an increasing number of tasks. Robotic grasping has long been a popu lar research topic, yet existing grasping algorithms still have room for improve ment. Challenges remain when robots need to grasp multiple objects or handle unknown objects that were not encountered during the training process. This dissertation proposes a simple algorithm that shifts the focus in deep learning from object classification to object features such as lines, curves, and corners. This approach enhances the generalization capability of the robot, allowing it to grasp a wider range of objects. Instead of relying on the commonly used method of manually annotating the best grasping point for each object, the pro posed algorithm enables the robot to automatically determine the optimal grasp ing point based on three types of object features. Additionally, through object segmentation and feature classification, the algorithm allows the robot to grasp multiple objects in a specific sequence with a high success rate, rather than being limited to individual objects. Experimental results demonstrate that, out of 275 grasping attempts, the algorithm achieved an 88% accuracy rate in de termining the optimal grasping point, with an actual grasping success rate of 85.1%.