Imitation Learning Based Task Completion with Drones

Imitation learning provides a novel method of teaching robots to perform tasks. In this thesis, previous successes in imitation learning are built upon by teaching a drone to perform a task that involves search and obstacle avoidance. The aim of the research, which is performed in a simulator, is to determine whether specific imitation learning methods can be used to complete the task. Example data is collected by an expert at the task and used to train two types of neural networks. An algorithm is then applied to introduce error recovery data into the dataset. Six models are evaluated using a combination of behavioural and statistical methods. The results indicate that while the models are not safe for application on a real drone, the method shows promise for future work.