Control and perception methods for agile flight of multi-rotor UAVs

Abstract

Multi-rotor Unmanned Aerial Vehicles (UAVs) have found extensive applications in our daily lives. Applications such as aerial surveillance or monitoring tasks benefit from the agile flight capabilities of UAVs. Agile flight allows UAVs to navigate complex environments, swiftly change flight trajectories, and capture images or videos from various angles. In critical search and rescue missions, an agile UAV can rapidly search a large area and respond to dynamic situations. Enhancing UAV agility requires a holistic approach involving factors such as advanced control systems and robust perception techniques.

Dynamics modeling plays a critical role in achieving precise control. Starting from dynamics modeling, this thesis firstly addresses the modeling and control of non-minimum phase dynamics of a bi-copter UAV. The presence of non-minimum phase dynamics poses challenges to hovering control, necessitating the design of a robust controller. To this end, first-principle model of dynamics is derived, identified, and subsequently employed in the design of an H$_\infty$ loop shaping controller. The experimental results demonstrate that the bi-copter is capable of hovering robustly using the designed controller. Then hybrid control for multi-mode flights are designed to make the bi-copter UAV achieve stable control under different situations and switch the working mode according to mission requirements.

In addition to achieving precise control, rapidly avoiding moving obstacles is crucial during agile flights in dynamic environments. Conventional approaches often divide the UAV system into distinct components such as localization, perception, planning, and control, which requires significant computation resources and leads to considerably high latency. In this thesis, a novel robocentric model-based visual servoing method is proposed, which uses the direct estimation of visual targets to establish a robocentric model of UAVs, enabling integrated state estimation, trajectory planning, and tracking control. The method promises higher bandwidth, and its effectiveness is verified through successful experiments in moving obstacle avoidance.

One of the critical challenges in avoiding moving obstacles is the timely detection of moving obstacles, which involves efficiently extracting obstacle information from abundant sensor feedback. To address this challenge, an efficient method for detecting moving objects using LiDAR, M-detector, is developed in this thesis. It is a system that leverages LiDAR point streams to detect moving events in the scene. By exploiting the occlusion principle, M-detector can detect moving objects accurately, immediately, and with high generalization ability. M-detector finds applications in various domains, including UAV dynamic obstacle avoidance, autonomous driving, and traffic monitoring. Furthermore, this thesis proposes a two-stage framework that combines M-detector with 3D object detectors to achieve object-level moving detection, which suggests the significant potential of M-detector.