Controlador baseado em aprendizado por reforço para robôs seguidores de pessoas

Abstract

This work proposes a tracking and follow-up technique training of people through follower robots, using learning by reinforcement. The approach is based on the Deep Q-Network (DQN) algorithm, which combines reinforcement learning and deep neural networks. The implementation ˜ consists of a customized training environment and an agent capable of learn to move around from a scenario created in the Gazebo simulator. ' The work presents a bibliographic review on the subject, followed by the description of the methodology used, which involves research in the databases, construction of the solution and training of the model. As a result, the technique is proposal proved to be feasible to be applied and tested in real robots and has potential for several applications, such as surveillance, search and rescue, as- human assistance and security.