Autonomous Aerial Recovery: Fixed-wing UAV Ballistic Airdrop and Deep-Stall Landing
MetadataShow full item record
Although human beings don’t have wings, it is often very useful to fly. Or at least to use the air to transport objects. Sometimes these objects have to be delivered on normally unreachable places under difficult conditions, and more often than not, it is valuable to be able to repeat this mission. This PhD thesis contributes to the research on autonomous fixed-wing unmanned aerial vehicles (UAVs) operating in difficult conditions. The aim of this thesis is for the UAV to perform a high-precision drop of an object and then return to base, landing smoothly on a small surface. To achieve this, two topics have been investigated: autonomous ballistic airdrop from a small UAV and deep-stall landing of a fixed-wing UAV using nonlinear model predictive control (NMPC)-based guidance. On the first topic, a fixed-wing UAV is responsible for the delivery of a generic object on a target on ground. Through iterative development, a closed-loop autonomous system running on-board the UAV is created. Using machine vision to identify the target, the system plans a release point for the object, guides the aircraft to the release point and delivers the object with high precision on the target. The system consists of three separate subsystems that are integrated in the DUNE unified navigation environment, and each subsystem has been tested through simulations and through flight tests. The first subsystem is the target recognition and position estimation subsystem. The second subsystem is the airdrop planning subsystem, which receives target position estimates as input, calculates an ideal release point considering height, wind and speed and plans a path to this release point. The third subsystem is the guidance and control subsystem, which is responsible for guiding the UAV to the release point. The second subsystem, the system integration and experiments are main contributions in this thesis. The system flight tests show results with average target error distance of less than C meters, from 50 meters altitude and 18 m/s speed, and high repeatability. On the second topic, the UAV performs a controlled deep-stall landing. In a deep-stall landing, we exploit the advantages of aerodynamics and can land the UAV with a steep flight path angle at the same time as the speed is lowered. A guidance system for the UAV is developed, where the UAV is landed at an appointed landing site with minimal speed. To achieve this, an NMPC is tailored for the problem. The NMPC uses a prediction model of the system to plan a series of control actions over a time horizon, which guide the UAV to the landing site in a deep-stall. After developing an NMPC for the deep-stall landing with the desktop computer tool CasADi, the controller is tested with 3 DoF and C DoF models without considering modelling errors. Then the controller migrates to the code-generated embedded version of the real-time optimization tool ACADO, to fit an embedded computer on-board a UAV. The code-generated guidance system is implemented in DUNE as well and tested in a software-in-the-loop (SITE) environment with autopilot software and an advanced simulator. The outcome of this is an autonomous landing system that plans a path to a deep-stall landing start point and guides the UAV through this path and in the landing phase. The results of the SITE tests show good precision and low landing speed for low wind speeds, but emphasize the importance of a correct prediction model.