Visual Perception System for Aerial Manipulation: Methods and Implementations

• Autores: Pablo Ramón Soria
• Directores de la Tesis: Aníbal Ollero Baturone (dir. tes.), Begoña C. Arrue Ullés (dir. tes.)
• Lectura: En la Universidad de Sevilla ( España ) en 2019
• Idioma: español
• Número de páginas: 169
• Tribunal Calificador de la Tesis: Emilio Gómez (presid.), José Guillermo Heredia Benot (secret.), Reinhard Braunstingl (voc.), Luis Antidio Viguria Jiménez (voc.), Alberto Vale (voc.)
• Programa de doctorado: Programa de Doctorado en Ingeniería Automática, Electrónica y de Telecomunicación por la Universidad de Sevilla
• Enlaces
  • Tesis en acceso abierto en: Idus
• Resumen

  • Technology is growing fast, and autonomous systems are becoming a reality. Companies are increasingly demanding robotized solutions to improve the efficiency of their operations. It is also the case for aerial robots. Their unique capability of moving freely in the space makes them suitable for many tasks that are tedious and even dangerous for human operators.

    Nowadays, the vast amount of sensors and commercial drones makes them highly appealing. However, it is still required a strong manual effort to customize the existing solutions to each particular task due to the number of possible environments, robot designs, and missions. Different vision algorithms, hardware devices, and sensor setups are usually designed by researchers to tackle specific tasks. Currently, aerial manipulation is being intensively studied to allow aerial robots to extend the number of applications. These could be inspection, maintenance, or even operating valves or other machines.

    This thesis presents an aerial manipulation system and a set of perception algorithms for the automation of aerial manipulation tasks. The complete design of the system is presented and modular frameworks are shown to facilitate the development of these kinds of operations.

    At first, the research about object analysis for manipulation and grasp planning considering different object models is presented. Depend on the model of the objects, different state of art grasping analysis are reviewed and planning algorithms for both single and dual manipulators are shown.

    Secondly, the development of perception algorithms for object detection and pose estimation are presented. These allow the system to identify any kind of objects in any scene and locate them to perform manipulation tasks. These algorithms produce the necessary information for the manipulation analysis described in the previous paragraph.

    Thirdly, it is presented how to use vision to localize the robot in the environment. At the same time, local maps are created which can be beneficial for the manipulation tasks. These maps are enhanced with semantic information from the perception algorithm mentioned above.

    At last, the thesis presents the development of the hardware of the aerial platform which includes the lightweight manipulators and the invention of a novel tool that allows the aerial robot to operate in contact with static objects.

    All the techniques presented in this thesis have been validated throughout extensive experimentation with real aerial robotic platforms.