Controle inteligente de um robô móvel utilizando modos deslizantes, redes neurais artificiais e aprendizagem por reforço

• Ano de defesa: 2021
• Autor(a) principal: Baumann, Gabriel de Albuquerque Barbosa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA MECÂNICA
• Palavras-chave em Português: Controle inteligente não linear; Robôs móveis; Controle por modos deslizantes; Redes neurais artificiais; Aprendizagem por reforço
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/48351

Research on intelligent and autonomous mobile robots has grown significantly due to its military, civil and industrial applications, such as the monitoring of agricultural plantations, the use in actions to support environmental disasters, border patrol, mapping of submarine territories or even the study of animal behavior. This work rescues the multi and interdisciplinary motivation of artificial intelligence, starting from philosophical questions to reach the characterization of intelligent and autonomous systems. Thus, only after building the theoretical bases for the concept of these agents, a bioinspired approach is presented for the trajectory tracking task by a omnidirectional mobile robot, the Robotino® produced by Festo® . For this purpose, the strategy consists of robust non-linear intelligent control using Sliding Modes, artificial neural networks and the Upper Confidence Bound algorithm. Each of these fundamentals techniques are presented, in order to justify, in advance, their consistent use with the theoretical proposal, to be later incorporated into the controller. Thus, Sliding Modes and their limitations regarding residual error are presented; artificial neural networks are then applied with the purpose of reducing them, however, they also have their restrictions; the Upper Confidence Bound is therefore added in order to mitigate them. The characteristics of each technique give the robot robustness in the control task, learning and autonomy with decision-making, respectively, as explained from the numerical and experimental results. The designed algorithm not only achieved the purposes, but also brought other positive points, such as avoiding the neural networks divergence resulting from the continuous updating of their weights. The approach developed based on the most recent arguments about autonomous agents obtained excellent results in both simulations and in experiments for the Robotino® trajectory tracking problem and represents the growing trend of research in embodied cognitive science.