Unmanned ground vehicle indoor navigation based on deep reinforcement learning

This dissertation aims to provide the methods of using Deep Reinforcement learning algorithm to train the UGV in simulation such that the trained UGV can reach a random target position and avoid the obstacles without any prior knowledge and model of environment. First, the basis of reinforcement learning, deep learning and deep reinforcement learning is introduced in chapter 2. In chapter 3, the the detail approaches used in this dissertation are described, including the software tools and algorithms that are used to build the simulation environment for training. We use three advanced and prevalent deep reinforcement learning algorithms to solve the expected tasks and design novel reward functions to increase the convergent capability. The Whole objective is divided into three steps, and the implementation process is included in chapter 4, where the main results are shown. The technical discussion and analysis about the problems of training the reinforcement learning system are included in chapter 5. Finally, the conclusions and recommendation to the future works are presented in chapter 6. Keywords: UGV, target reaching, obstacles avoidance, deep reinforcement learning, reward function.Master of Science (Signal Processing