Robot Grasping Based on RGB Object and Grasp Detection Using Deep Learning

Abstract

In the last decade, great emphasis was given to the implementation of robots in the industry, but they are still applying traditional techniques. Therefore, industrial plants with continuous production come to work without any type of feedback, which causes uncertainty and possible incidents. The present project shows the development of a system for the detection, recognition, and grasping of objects using a robotic arm. The system was developed with SSDResNet50 architecture for object detection and in the case of the grasping area detection, we propose an architecture that deploys a grasping area represented on a vector of 4 values. On the other hand, we made the integration and simulation of all developed sub-systems to perform grasping on Robot Operating System (ROS). The system achieves a minimum error in the detection of objects, reaching a value of 94.0% in the evaluation mAP, for the detection of the grasp area, a 26.32% is obtained in the Jaccard metric. It should be noted that a dataset generated within the simulation environment has been used. Finally, the grasping success of each object was evaluated, obtaining an average rate of success of 85.76% among all objects.