Desarrollo de un Algoritmo de Navegación Autónoma Basado en Técnicas de Aprendizaje por Refuerzo Usando Información Visual

Abstract

This project proposes the implementation of an algorithm autonomous navigation based on visual information using deep reinforcement learning. The algorithm aims to teach an agent to identify visual patterns to navigate to a goal in closed and unknown environments. The learning process is made out of three stages: Classification, Imitation and Training, and a Replay Memory system. The Learning stages provide the agent with different tools to classify the information and make a decision, transferring the knowledge acquired in each one. Meanwhile, the replay memory provides the agent information from past experiences to understand and solve unfamiliar environments. At the same time, the algorithm is based on a Deep Q Network (DQN) model, with a reward to the agent in each interaction with the environment. The evaluation of the algorithm is performed through experiments based on the interaction with simulated environments of different sizes, routes and features.