The combination of reinforcement learning and deep neural networks has the potential to train intelligent autonomous agents on high dimensional sensory inputs, with applications in flight control. However, the amount of samples needed by these methods is often too large to use real-world interaction. In this work, mirror-descent guided policy search is identified as a promising algorithm to train high-dimensional policies on real-world samples. Several experiments are conducted to investigate how the use of expert-demonstrations can further improve the sample-efficiency of this algorithm when applied to the control of a quadcopter in simulation. It is shown how demonstrations, when combined with certain alterations in the mirror descent guided policy search algorithm, can significantly reduce the amount of samples needed to achieve good performance. Additionally, it is shown how these improvements are robust to sub-optimal demonstrations.

The urge for adventure combined with the global need for increasingly sustainable methods of air transportation resulted in a pioneering idea. What better way to show the possibilities of solar power by altering the accessible and entertaining nature of a paraglider? Personal experience and vision inspired the client to be the first to fly such an aircraft over the Atlantic Ocean to spread awareness for a more sustainable future. The mission was born: “Fly manned, continuously across the Atlantic using a solar powered paraglider.” Crossing the Atlantic requires the typical range of a paraglider to drastically increase. This combined with the current technology in terms of energy storage and solar power generation makes designing a struggle. Duration is a problem due to the slow, inefficient cruise phase. The purpose of this report is to investigate the feasibility of flying a solar powered paraglider over the Atlantic. The flight route and time of execution were based on the least amount of energy required to make the crossing. This means flying the shortest distance, with the most amount of sun hours per day, the highest solar intensity and the most advantageous tail winds. Impossible to maximise every parameter an optimal mission was determined to fly a rough 3040 km from St. John’s, Newfoundland Quebec to Portmagee, Ireland during the late June summer days. The design mainly evolved around flying with the least amount of energy required whilst still allowing a duration short enough for the pilot to cross without falling asleep. Flying efficiently typically means flying slow, but the aforementioned duration set a lower bound to this. The maximum mission duration was extended as much as possible by techniques such as blue light, stimulating substances, pilot health and pilot entertainment. By designing every part of the paraglider to be as lightweight as possible results in a cycle of less energy required, which in turn allows smaller subsystems, which in turn has lower weight. The subsystems were designed in a concurrent way. Progress in one part means changed inputs for another part. Integrated MATLABr tools based on analytic models were created to converge to an optimal design point for the subsystems in relation to each other. The subsystems and final product were analysed for sensitivity of their input values which helped identify the best solution for improvements. The final product was also analysed for its performance in order to check requirement compliance and the actual mission feasibility. The missions two most important limiting factors are the mission duration and the power required. In order to minimise mission duration the flight speed should increase whilst the opposite is true for the power required. This study revealed that power required was the limiting factor so a lower flight speed was chosen. For the chosen cruise speed of 10 m· s−1 the resulting mission duration was 38 hours and the power required ended up to be 3.79 kW. The total mass of the vehicle is 277 kg of which 23% is battery mass and 16% solar cell mass. The high energy density lithium-sulphur batteries are still in development, but are expected to be available within 5 years. Currently available batteries would make the design too heavy. This study was initiated as a concept study. With a custom designed wing and fuselage the mission is shown to be technically feasible in 5 years time. The limited scope and duration of this project allows for more improvements for which this report can be used as a guideline.