A simulation-based machine learning approach for flight control system design of agile maneuvering multicopters

Ability to do agile maneuvers provides an enhanced autonomy level for operations in dense and dynamically changing environments involving urban or nap-of-the-earth flights. In this work, we consider the design of agile maneuvering flight control system through a hybrid machine learning environment which includes both real flight testing and simulation-based deep learning. The main aim is to be able to identify the highly non-linear dynamical characteristics and control patterns that a pilot exploits (in many cases as second-nature) and utilizes while performing agile maneuvers. Specifically, the in-house developed high performance multicopters and their models are used for performing agile maneuvers under pilot control. Later this data is used in designing agile maneuvering outer loop controller designs via deep learning as to mimic (and even enhance) pilot’s agile maneuvering capability on a quadrotor. The control design that we consider is a cascaded flight control system design which consists of classical attitude control loops and neural network outer loops. In the first use case, we considered the pilot controls the velocity of the quadrotor by using a classical attitude control system and input/output data is collected on a data logger. Then, the collected data is used to train a neural network based velocity controller and system performance is evaluated in a test case. In the second use case, we considered, lazy eight maneuver and racetrack patterns are built in the Gazebo environment, which contains several gates. The pilot performs manual flights in this patterns and tries to fly through the defined gates. In these flights, required input/output dataset is collected and it is used to train a neural network-based guidance algorithm, which is developed for autonomous maneuvers. In the third use case, we considered a new racetrack pattern is built within the Gazebo environment, which is not used in the data collecting process of the second use case. Then, the proposed neural network based guidance algorithm is evaluated in the new racetrack pattern without performing any data collecting and training process. The results show that a) the quadrotor can even do maneuvers for which does not specifically exist within its training set, and b) the quadrotor can perform autonomous agile maneuvers at high speeds that are not demonstrated by classical control methods.