Flying through Gates using a Behavioral Cloning Approach

Abstract

Drone racing presents a challenge to autonomous micro aerial vehicles (MAV) because usually the track is not known in advance and it is affected by the environment light. In such scenarios, the vehicle has to act quickly depending on the information provided by its sensors. In this work, we want to predict the movement of the drone so that it passes through a gate. Unlike previous approaches where the task is decomposed into perception, estimation, planning, and control, we are proposing a behavioral cloning approach. In this method, a convolutional neural network is trained with the flights of a human operator. So that the output of the trained network is directly the desired MAV state so that it leads the drone through the gate. We have tested the method using a validation set where we obtained a low loss. Furthermore, we have tested the trained network with unseen data obtaining promising results.