Implementação de um Sistema Autônomo de Construção de Estrutura usando Aprendizado por Reforço

• Ano de defesa: 2017
• Autor(a) principal: DANTAS, Diego de Oliveira
• Orientador(a): SANTOS, Sérgio Ronaldo Barros dos
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO/CCET
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: Robótica; Aprendizado por reforço; Learning Automata; Construção autônoma; Robotics; Reinforcement learning; Autonomous construction
• Área do conhecimento CNPq: Engenharia de Software
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/2062

This work presents the development and implementation of an autonomous construction system in which uses a terrestrial mobile robot for constructing three-dimensional structures from blocks of different size. A high level planning is proposed to generate the construction plans of the structures. This algorithm is based on Reinforcement Learning methods called Finite Action-Set Learning Automata (FALA) and Parameterized Learnig Automata (PLA). From this planner, the used types of blocks for the construction and the final composition of the structure is defined by the user. The high level planner is used to solve the following problems: 1) Generate an optimal assembly diagram, which consists of a list of positions, orientations and kind of blocks final, taken into account the design of the structure defined by the user. The minimal number of blocks and also the restriction of assembly is considered during the generation of the diagram; 2) Generate an optimal execution plan that can be used by the robot to accomplish the task of assembly. This plan is composed by the sequence of procedures for manipulating and assembling blocks. The trajectories generated by the global planner based on A* algorithm is used to accomplish the execution plan. After completion the execution plan, the global planner sends a series of positions to a path tracking controller, called eband local planner. This tracking controller is used to control the robot during it navigation through simulated or actual environment. The mapping of the simulated and real environments and the location of the robot in the environment is performed using the algorithm called Real-Time Appearance-Based Mapping (RTAB-Map). The RTAB-Map uses image and odometry information to generate the environment also estimate the position of the robot in relation to the global coordinate system. The simulated and actual robots use the framework called Robot Operation System (ROS). The ROS allows the communication between different applications even if they are performed in different machines. To demonstrate the efficiency of the obtained solutions using the high level planner, simulated and experimental tests of the autonomous construction system are performed. During these tests, different types of structure (tower, containment wall, space station and pyramid) are assembled. The results show that the reinforcement learning method is able to feasible assembly diagrams and execution plans (sequence of procedures) can be used to perform the task in a short period of time.