Robot kontrol with deep reinforcement learning in simulation environment

Reinforcement Learning, a branch of machine learning, is a method of learning based on choosing actions to maximize rewards in an environment. This form of behavioral learning, which is widely observed in nature, gives successful results for autonomous tasks in environments where there is no information about. We can train mobile robots for autonomous tasks with Deep Reinforcement Learning, one of the Reinforcement Learning methods. Mobile robots can perform path and movement planning tasks by mapping the environment with simultaneous positioning and mapping (SLAM) algorithms. However, operating in uncharted environments is a challenging task. In our study, we aimed to overcome these difficulties by using Deep Reinforcement Learning methods to perform specified tasks in unknown environments with mobile robots. The necessary environment was created with the Robot Operating System and the Gazebo simulation environment, which can work with the 3D models frequently used on simulation platform, and the TurtleBot3 mobile robot aimed to reach the determined goals in this environment without hitting obstacles. Due to the limitations of Q Learning, the frequently used method of reinforcement learning, Deep Q Networks, one of the Deep Reinforcement Learning methods that use neural networks instead of memory for complex problems, were preferred in our study. Training was conducted with the TurtleBot3 mobile robot in the environment formed in Gazebo. According to the greedy approach policy, which is one of the basic features of reinforcement learning, a model that can find the target point many times and increase the reward it collects throughout the episode, by leaving the movements made randomly at the beginning to the movements decided by the neural network model in the later stages, has been constituted and performance of this model has been showed with graphics.