Adaptive non-Linear Model Predictive Control for UAV-UGV Coordination

Unmanned Aerial Vehicles (UAV) have gained a great amount of popularity in the last years. Among the features that make them so successful are the cost, the portability, their airborne nature and their ability to perform Vertical Take-Off and Landing (VTOL). Since the different tasks that UAVs are given may require them to travel long distances, the small batteries that they are using can prove to be a problem if they have to return to the home-base to recharge. A solution can be sending a Unmanned Ground Vehicle (UGV) along with the UAVs, since they can have a much longer autonomy and provide a moving base. In particular this thesis wants to improve the landing manoeuvre in such a way that the landing will be smooth and most accurate. The first part of this work is dedicated the implementation of the Non-linear Model Predictive Control (NMPC) as the baseline controller. This controller creates an optimal control that minimizes a given cost function along the chosen prediction horizon, while respecting the model constraints. For this work the non-linear version of the algorithm was chosen given the non-linearity of the model of the UAV in question, i.e. the quadrotor. One of the main features of this controller is the dependency on the internal model of the system, which gives it its name. In the second part another complementary controller is added to the NMPC. Adaptive control is meant to deal with varying or unknown model parameters and for this reason L 1 Adaptive Control (L1AC) was chosen for the task. L1AC is a slight variation of the Model Reference Adaptive Control (MRAC), which tries to deal with errors at the input level, setting a new input counterpart to negate them. The combination of the two controllers is meant to exploit each other’s strength, while helping with the other’s weakness, to create an overall stable and well performing algorithm that allows to account for the kinematics and dynamics of the quadrotor, while being extremely robust against the potential errors of the internal model or external disturbances. The proposed method is then tested in various scenarios and validated through results on Matlab & Simulink.