Learning Autonomous Quadcopter Trajectories from Demonstration

How to make a drone fly autonomously is a fascinating problem. This is an important field of research nowadays due to the rising usage of drones in numerous different application. Automatizing it would make them do their work autonomously and it can be very useful, for instance for security control or delivery, but this are just two example of their use. The aim of this thesis is to analyze this problem starting from a drone that can be piloted by a human giving it destination coordinates and find a method to make it fly autonomously. In this thesis this problem has been approached applying an Inverse Reinforcement Learning algorithm. Thank to this, after a training done with expert trajectories the drone started to fly alone given a starting position and a goal. During the development of the project dynamic obstacle have been used is order to make the environment closer to the reality. In the first chapters there is a description of some works previosly done on similar topic and a description of everything used during the research. Then there is a chapter where the implementation procedure is described. Here a step by step resolution of the problem is given to the reader. Then there is chapter which contains the results of the tests. They are described and explained. In conclusion the drone learned a cost function from observed trajectory and is able to compute the best path between two point in the map. In the case it sees an obstacle which is not in the map it adds this and change its trajectory in the case there is a risk of collision.