Uma abordagem baseada em Engenharia Dirigida por Modelos e Aprendizado de Máquina Aplicado a Robôs Móveis

Abstract

The application of mobile robots in complex environments requires a high capacity for autonomy in their decision making. The literature states that the next step in the evolution of autonomous robotic controllers is to make robots self-adaptive. In addition, advances in the field of Machine Learning are increasing, contributing to the emergence of numerous opportunities for the development of intelligent controls applied to robots. However, there are still several challenges to be faced, for example, the complexity in development, the need for reprogramming when changing aspects of the environment, the need for large amounts of data, error propagation, mapping complex states into a single robot, among others. In order to mitigate these problems, we propose a Framework, called RLoRDE (Reinforcement Learning for Robotic model-Driven Engineering), which is based on the use of Model Driven Engineering to simplify the development of robotic software, where the code is generated according to the model created by the developer following rules imposed by the metamodels developed during this thesis. A graphical tool assists in creating and transforming models to code. Reinforcement Learning methods are available where it is possible to generate training environments for missions that require flexibility to deal with the variability of the environment and promote self-adaptation. Our experiments were carried out by increasing the degree of complexity of the environment for the robot's mission. The experimental results show that the Framework RLoRDE is promising in the sense that we obtained an average 69\% mission success rate in scenarios where the robot was not trained.