This Dissertation addresses the design and development of small-scale UnmannedAerial Vehicles of the TiltRotor class, alongside their autonomous navigation requirements,including the fully-onboard state estimation, high-efficiency flight control,and advanced environment perception.Starting with an educated Computer Assisted Design-based methodology, a mechanicallyrobust, customizable, and repeatable vehicle build is achieved, relyingon high-quality Commercially Available Off-The Shelf equipment –sensors, actuators,structural components–, optionally aided by Rapid Prototyping technology.A high-fidelity modeling process is conducted, incorporating the rigid-body dynamics,aerodynamics, and the actuation subsystem dynamics, exploiting fistprincipleapproaches, Frequency Domain System Identification, as well as computationaltools. Considering the most significant phenomena captured in thisprocess, a more simplified PieceWise Affine system model representation is developedfor control purposes –which however incorporates complexities such as flight(state) envelope-associated aerodynamics, the differentiated effects of the directthrust-vectoring (rotor-tilting) and the underactuated (body-pitching) actuationauthorities, as well as their interferences through rigid-body coupling–.Despite the switching system dynamics, and –as thoroughly elaborated– theirreliance on constrained manipulated variables, to maintain a meaningful controlorientedrepresentation, the real-time optimal flight control of the TiltRotor vehicleis achieved relying on a Receding Horizon methodology, and more specifically anexplicit Model Predictive Control framework. This synthesis guarantees globalstability of the switching dynamics, observance of state and control input constraints,response optimality, as well as efficient execution on low computationa power modules due to its explicit representation. Accompanied by a proper Lowand-Mid-LevelControl synthesis, this scheme provides exceptional flight handlingqualities to the aerial vehicle, particularly in the areas of aggressive maneuveringand high-accuracy trajectory tracking.Moreover, the utility of TiltRotor vehicles in the field of aerial robotic forcefulphysical interaction is researched. Exploiting the previously noted properties ofthe PieceWise Affine systems Model Predictive Control strategy, the guaranteedstabilityFree-Flight to Physical-Interaction switching of the system is achieved,effectively bringing the aerial vehicle into safe, controlled physical contact withthe surface of structures in the environment.More importantly, employing rotor-tilting actuation –collectively and differentially–significant forces and moments can be applied onto the environment, while via thestandard underactuated authority the vehicle maintains a stable hovering-attitudepose, where the system’s disturbance rejection properties are maximized. Overall,the complete control framework enables coming into physical contact with environmentstructures, and manipulating the enacted forces and moments. Exploitingsuch capabilities the TiltRotor is used to achieve the execution of physicallydemandingwork-tasks (surface-grinding) and the manipulation of realisticallysizedobjects (of twice its own mass) via pushing.Additionally, the fully-onboard state estimation problem is tackled by implementingdata fusion of measurements derived from inertial sensors and customdevelopedcomputer vision algorithms which employ Homography and OpticalFlow calculation. With a proper sensorial setup, high-rate and robust ego-motionestimation is achieved, enabling the controlled aggressive maneuverability withoutreliance on external equipment, such as motion capture systems or GlobalPositioning System coverage.Finally, a hardware/software framework is developed which adds advanced autonomousperception and navigation capabilities to small-scale unmanned vehicles,employing stereo vision and integrating state-of-the art solutions for incrementalenvironment building, dense reconstruction and mapping, and point-to-pointcollision-free navigation. Within this framework, algorithms which enable the detection,segmentation, (re-)localization, and mobile tracking –and avoidance– of adynamic subject within the aerial vehicle’s operating space are developed, substantiallyincreasing the operational potential of autonomous aircraft within dynamicenvironments and/or dynamically evolving missions.
Deep Reinforcement Learning for End-to-End Local Motion Planning of Autonomous Aerial Robots in Unknown Outdoor Environments: Real-Time Flight Experiments