Autonomous navigation of differential robot with depth camera and laser range finder using Deep Reinforcement Learning in non-stationary indoor environments

Abstract

Intelligent systems are an area of computer science that seeks to solve complex and multidisciplinary problems by using an automatic focus by capturing information from which it can perform actions and evaluate its result in order to learn from its experience and improve its performance and efficiency. These systems can be applied in robotics for autonomous navigation purposes, where the use of conventional automatic control techniques is very complex. The topics with the greatest potential are Reinforcement Learning and Deep Learning, which through their combination is possible to solve highly complex problems from the training of computational frameworks based on rewards for taking a correct action or penalties for incorrect actions. The thesis project is focused on providing the autonomous navigation capacity to a differential robot making use of a Lidar depth camera and Deep Reinforcement Learning (DRL) techniques for use in indoor environments with non-stationary elements. For its development, a review of the DRL frameworks used for autonomous navigation will be carried out, from which the development of frameworks based on DRL techniques for autonomous robot navigation will be considered in indoor environments where elements are non-stationary. Subsequently, the evaluation of the frameworks in virtual platforms is proposed to determine its performance. Finally, a robotic platform with differential traction and a Lidar depth camera will be conditioned to be used in a validation stage through experiments in real environments.