Learning Model Predictive Control for Optimal Path Planning of Quadrotors in Multi-Scenario Applications

The aim of this thesis is to develop a Learning Model Predictive Control (LMPC) framework for quadrotors; this work results to be one of the first comprehensive studies on LMPC applied to quadrotors in real-case scenarios. LMPC is a novel control technique that autonomously learns to improve its performances by collecting data coming from past executions of the control task. Given a certain task, the control algorithm makes the quadrotor to perform it repetitively; information coming from these repetitive iterations is progressively collected and employed by the controller to obtain better performances for the required task. This control method is especially versatile and useful for time-sensitive operations, in which the drone has to make fast and dexterous maneuvers in constrained and cluttered environments. In this scenario, our specific goal is to implement a LMPC algorithm that pilots the quadrotor within a closed 3D race track, in which multiple types of obstacles can be inserted. The task of the controller is to autonomously find the trajectory achieving the minimum lap time, after multiple flights of the drone within the track. The control algorithm is fully developed in Matlab and is tested via several software-in-the-loop simulations, employing a complete dynamic model of the quadrotor. The conducted simulations show that the LMPC algorithm successfully achieves the task of finding the optimal path for lap time minimization and also the additional task of avoiding the obstacles placed within the track: the control algorithm has learned to fly the quadrotor aggressively, adopting a flight style that exploits multiple driving tricks to optimize both the travelled distance and the time needed to complete a lap. The simulations not only demonstrate the correct functionality of the algorithm, but also empirically verify all the fundamental theorems of LMPC.