Developing object grasping skills with deep reinforcement learningalgorithms

With the development of Artificial Intelligence in different areas, the increase in the use of robots in different application areas has also increased the need for interaction with their environment. For this purpose, various sensing components such as touch and vision have been used recently, while robot end effectors provide interaction with the environment. The ability to detect changes in the environment by the camera without considering the parameters of the joints of a robotic end effector has made visual detection tools such as a camera a very useful tool. In a vision-based system, it is important to determine the movements of the robot arm to fulfill a given task such as grasping with the environment properties taken from the images. In this thesis, a comprehension scenario was created in the Webots simulator environment. Based on the images taken from the camera, a basic robotics application framework is presented, in which it is aimed to measure the grip performance of the KUKA youBot robot arm with 5 degrees of freedom. The act of grasping was developed with a learning strategy based on deep reinforcement learning. In this context, firstly, a one-dimensional deep Convolutional Neural Network (CNN) structure was used in the automatic classification of grasp types in humans consisting of Surface Electromyogram signals. Later, a Deep Convolutional Q Network (DCQN) model was developed that can play the Pong game. Based on this developed model, a duelling DCQN method, which is a basic learning structure that can learn to grasp in the Webots simulation environment, has been successfully created with the KUKA youBot robot arm. As a result, a model of a developable grasping system supported by artificial intelligence was obtained in the simulation environment. With this thesis, object grasping skills encountered in many applications have been developed and in this direction, different algorithms based on deep learning and reinforcement learning are presented. In the thesis, after identifying the existing problems in the industry, classification and estimation skills for object grasping were developed. Especially in some processes of the industrial area, solutions to the problems of the users are presented.