CNN - Based Object Detection for Robot Grasping in Cluttered Environment

The objective of robot vision is to identify objects of interest within the robot’s workspace from digital images. Despite advances in the programming and control of collaborative robots, object recognition and localization remain a challenging task. Given that the performance of collaborative robots operating in cluttered environments is heavily reliant on robot vision, the implementation of an appropriate object recognition algorithm can significantly improve its performance. In this paper, a robot vision algorithm is developed based on a Convolutional Neural Network (CNN) classifier for object detection and spatial localization for robot arm manipulation and grasping and tested in cluttered scene in terms of its robustness to variations in light conditions, viewing angles, and other uncertainties and noise. After successful object recognition, homography is employed for object localization, which serves as input for the robot’s manipulation control. The experimental setup involves a collaborative robot equipped with a haptic gripper and a static camera placed above the manipulation area. The results of the robot’s manipulation and grasping based on the implementation of developed algorithms are presented and discussed in final chapters of the paper.