AUTONOMOUS MOBILE ROBOT NAVIGATION DESIGN USING DEEP REINFORCEMENT LEARNING

• Main Author: Izaty, Annisa
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/74702
• Institution: Institut Teknologi Bandung

Mobile robots are categorized as autonomous when they have the ability to well- navigate so that they can move from the starting point to the target without colliding the obstacles. A path planning system is needed to fulfill the task. Therefore, the objective of this Thesis is to generate a path planning system, globally and locally, which optimizes the reached distance, time, and efficiency of memory usage. In the global path planning, We use the algorithm which is Rapidly-Exploring Random Trees (RRT*) combined with fixed node, mesh grid approaches and visibility graph sampling method. Meanwhile in local path planning, Dynamic Window Approach (DWA) is combined with Deep Reinforcement Learning (DRL) to determine the optimal translation velocity and angular velocity. Then, the robot will execute the global path using local path planning. DRL algorithm which is already trained in Gazebo ROS determine the pairs of velocities (v,?). We carry out the training in simulation environment and real environment with 3 circumstances, which are without obstacles and with agent training in the first environment; with obstacles and agent training in the second environment; and with obstacles but without agent training in the third environment. In this research, we execute 1000 episodes of training. From testing process, in the first environment, the success rate of the robot to reach the target in the simulation environment is 100%, meanwhile in the real environment is 80%. In addition, in the second environments, the success rate for simulation and real environment is 90% and 60%, while in the the third environment is 80% and 60%, respectively. Those results lead to the conclusion that robot is able to be well-trained in two types of environments, and also it is able to reach the target from the starting point in all 3 types of environments with the shortest distance, time, and efficient usage of computation memory. However, to achieve more optimum results, it is needed to carry out a training with a greater variations of environment and number of training episodes.