Robust algorithms for the identification and control of Android-powered quadcopters

The focus of this thesis is on modeling and control of a non-real time, Android operated quadcopter of type "dji F450 Flame Wheel" without having a concrete knowledge about the system's dynamics and parameters. The quadcopter is equipped with an onboard non-rooted Android smartphone, which serves as both the controller and the IMU in the system. The reference command signals are generated by another user-held Android smartphone which defines the desired orientation of the quadcopter. Due to the fact that default Android implementation is not real-time, the measurements of both Android phones are subject to significant latencies resulting in asynchronous data. To obtain a model of the system, a comprehensive system identification study of the quadcopter's rotation dynamics using grey box model and Euler's equations of rigid body is introduced in the thesis. It also introduces two novel algorithms for obtaining an initial guess for the inertia matrix using convex optimization despite the presence of large number of local minimizers in the original prediction error problem. It shows how sensitive the process is to the initial guess of the model's parameters. A detailed comparison of the relevant estimates is also shown. The control laws were implemented on the onboard Android device, which reads the asynchronous built-in sensors measurements and generates the control signals required to steer the quadcopter and obtain the desired orientation defined by the user-held device. Two control laws were developed, an advanced model-free PID controller that accounts for the non-uniformly distributed data, the windup effect, and the derivative kick, and a model-based LQI controller. Both control approaches were able to stabilize the quadcopter despite the data asynchronousity and model uncertainty, and were validated and tested empirically and through simulation. The thesis also introduces a novel approach of optimizing the PID controllers gains based on the jacobian matrix. The optimization problem tends to be poorly conditioned for such systems. Hence, the novel scaling technique improves the conditioning of the optimization problem and obtains better minimizers. The efficiency of the proposed algorithm is evaluated through simulation. Furthermore, a detailed study on the effect of the cost function selection and model uncertainty on the optimization process is shown.