Aplicação de aprendizado por reforço em navegação de rôbos

Abstract

The study and the usage of robots to assist humanity has been studied deeply since the past century. One of the main researches is to perform the robot motion autonomously, safely and efficiently, in such a way that they perform tasks that may need locomotion. However, it is common that the desired path be complicated to build or follow, while some constraints of the environment have to be considered, such as, obstacle avoidance, moviment constraints or limitation on robot sensors. Recently, one of the areas that has achieved notoriety in the research community is deep reinforcement learning, which assembles concepts of reinforcement learning, one sub-area of machine learning, with the lastest breakthroughs of deep learning, another research field with several expressive results. Even considering that the first applications were in video games, many researchers have been proposing to apply these techniques in robot systems, for many tasks, for example, manipulation and locomotion. In this way, this dissertation will present some tools and algorithms recently proposed in deep reinforcement learning, which will make the robot capable of learning to move to a target in a scenario with obstacles. Besides that, this work will propose an algorithm that performs the learning of the best path according to the task continuously, improving the path travelled as the robot finalizes the tasks.