Resumen de Guidance, navigation and control of multirotors
This thesis presents contributions to the Guidance, Navigation and Control (GNC) systems for multirotor vehicles by applying and developing diverse control techniques and machine learning theory with innovative results. The aim of the thesis is to obtain a GNC system able to make the vehicle follow predefined paths while avoiding obstacles in the vehicle's route. The system must be adaptable to different paths, situations and missions, reducing the tuning effort and parametrisation of the proposed approaches. The multirotor platform, formed by the Asctec Hummingbird quadrotor vehicle, is studied and described in detail. A complete mathematical model is obtained and a freely available and open simulation platform is built. Furthermore, an autopilot controller is designed and implemented in the real platform. The control part is focused on the path following problem. That is, following a predefined path in space without any time constraint. Diverse control-oriented and geometrical algorithms are studied, implemented and compared. Then, the geometrical algorithms are improved by obtaining adaptive approaches that do not need any parameter tuning. The adaptive geometrical approaches are developed by means of Neural Networks. To end up, a deep reinforcement learning approach is developed to solve the path following problem. This approach implements the Deep Deterministic Policy Gradient algorithm. The resulting approach is trained in a realistic multirotor simulator and tested in real experiments with success. The proposed approach is able to accurately follow a path while adapting the vehicle's velocity depending on the path's shape. In the navigation part, an obstacle detection system based on the use of a LIDAR sensor is implemented. A model of the sensor is derived and included in the simulator. Moreover, an approach for treating the sensor data to eliminate the possible ground detections is developed. The guidance part is focused on the reactive path planning problem. That is, a path planning algorithm that is able to re-plan the trajectory online if an unexpected event, such as detecting an obstacle in the vehicle's route, occurs. A deep reinforcement learning approach for the reactive obstacle avoidance problem is developed. This approach implements the Deep Deterministic Policy Gradient algorithm. The developed deep reinforcement learning agent is trained and tested in the realistic simulation platform. This agent is combined with the path following agent and the rest of the elements developed in the thesis obtaining a GNC system that is able to follow different types of paths while avoiding obstacle in the vehicle's route.
