Neural Drone Racer Mentored by Classical Controllers

The autonomous drone race has driven the development of various approaches to agile flight involving perception, planning and control. The latter is most effective when using external information such as high-frequency drone position. On the other hand, deep learning-based methods propose to train an artificial pilot to learn to associate sensory data with flight commands using only visual information. In this paper, we propose an approach for developing a neural drone racer based on the mentoring of three classical flight controllers combined with visual information. We performed a comparison of the three trained models. Our results demonstrated an improvement in navigation behaviour and increased the speed to complete a racetrack by a factor of 7 compared to that reported in the state of the art.