Learning based methods applied to the MAV control problem

Graduation date: 2010This thesis addresses Micro Aerial Vehicle (MAV) control by leveraging learning based techniques to improve robustness of the control system. Applying classical control methods to MAVs is a difficult process due to the complexity of the control laws with fast and highly non-linear dynamics. These methods are mostly based on models that are diffcult to obtain for dynamic and stochastic environments. Due to their size, MAVs are affected by wind gusts and perturbations that push the limits of model based controllers where the linear approximation no longer holds. Instead, we focus on a control strategy that learns to map MAV states (e.g., heading, altitude, velocity) to MAV actions (e.g., actuator positions) to achieve good performance (e.g., flight time, minimal altitude and heading error) by maximizing an objective function. The main difficulty with this approach is defining the objective function and tuning the\ud learning parameters to achieve the desired results. These learning based techniques have been used with great success in many domains with similar dynamics and are shown to improve MAV robustness with respect to wind gusts, perturbations, and actuator failure. Our results show significant improvements in\ud response times to minor altitude and heading corrections over a traditional PID controller. In addition, we show that the MAV response to maintaining altitude in the presence of wind gusts improves by a factor of five. Similarly, we show that the MAV response to maintaining heading in the presence of turbulence improves by factors of three. Finally, we show significant improvements in the case of control surface actuator failure when using a multiagent system. The multiagent control system performs up to 8 times better than the PID controller when\ud tracking a target heading