Mobile robots: a study on sensing and perception systems

Abstract

Mobile robots are equipment used to perform tasks that require the equipment to move around the environment. Mobile robots are equipment used to perform tasks that require the equipment to move around the environment. Mobile robots can use machine learning techniques to perform intelligent tasks, such as recognizing objects and making decisions. For a robot to make decisions, it needs to collect data from the environment, process it, and convert it into information. This thesis aims to study environment sensing techniques traditionally used by mobile robots and to propose a new sensing technique, which uses the data provided by the sensors, processes them, and returns information. To achieve the objective of the thesis, a study was carried out on the traditional sensors used in mobile robotics and, then, a new sensing approach was proposed. First, a specific mapping approach was developed for Liquefied Petroleum Gas (LPG) storage tanks só that a climbing inspection robot, developed by UTFPR in partnership with Petrobras, was able to carry out preventive inspection on LPG tanks. This technique could predict the entire surface of the environment without the need for a complete scan. Subsequently, the application of intelligent 3D data processing techniques for navigation and self-preservation of unmanned aerial vehicles (UAVs) was studied. A navigation technique was developed in formation for 4 UAVs, avoiding collisions with the environment and between them, always maintaining the formation throughout the route. This action was only possible by processing the 3D sensor data, converting it into distance information from the center of the UAV, and performing obstacle avoidance tasks and self-preservation. From these first two works, the need to process the sensors’ data to generate useful information for robot decision-making became evident. The next paper of the thesis aimed to develop an approach for processing data from 3D sensors and RGB images to generate information, which can be used by a robot. The approach consisted of using computer vision to identify objects in an RGB image and point cloud processing to identify these objects in the real world. Subsequently, the approach was embedded in a compact device, called a DeepSpatial sensor. This equipment was coupled to a robot and validated for traditional applications in mobile robots, proving the sensor’s information’s efficiency. As a result of this thesis, a new sensing approach was proposed, where traditional sensors are used for intelligent actions. The approach is embedded in a compact device, which can be considered as a new sensor.