基於分散式非線性模型預測控制之多無人機障礙環境可形變編隊追蹤控制

Abstract

本論文旨在針對多無人機系統於障礙物空間內之可形變編隊追蹤問題進行研究，有別於傳統單一理想隊形之編隊追蹤問題，我們提出一創新之隊形表示法，其得以使理想隊形依據當下需求做出必要且合理的調變。為了避免多代理人系統之潛在數量級限制，本研究將採用相較於集中式通訊更具挑戰性之分散式通訊架構，機間僅傳遞部分且必要資訊以降低冗餘頻寬耗損。本論文之控制目標有三:其一，給定一理想隊形組合，由於可能受到環境或障礙物之影響而無法完全實現，此時系統將會按給定之形變標準為當下環境計算出較適合理想隊形。其二，透過分散式之通訊架構，依僚機間相對位置關係實時保持給定隊形。其三，為所有於障礙空間內執行編隊追蹤任務之代理人計算出無碰撞軌跡。對於以上控制目標，我們提出一基於主從關係(leader-follower)之分散式非線性模型預測控制(DNMPC)架構。在本架構下長機(leader)將負責計算當下適配之隊形並傳遞全域追蹤軌跡予以其他僚機(follower)，即控制目標一。其餘控制目標將以目標函數(objective function)與限制關係(constraint)的方式整合入我們提出之控制框架內，其中針對隊形維持之目標函數將透過實時僚機資訊盡可能在飛行過程保持理想隊形，而障礙物則被視為一限制關係以確保所有計算之軌跡皆無碰撞可能。透過我們提出之DNMPC架構，機群將得以順利於障礙空間穿行並依需求維持理想隊形之剛性程度，且得以透過可形變編隊設計在狹隘地形運算出較為可行的隊形編排。在本文末，我們亦提供一系列模擬與實際飛行試驗以驗證系統之效度。

The thesis considers multiple unmanned aerial vehicles (UAVs) deformable formation tracking control in an environment with obstacles. Unlike traditional formation tracking control problems with only a single desired formation, we aim to introduce a novel formation representation that can allow the necessary deformation between two desired patterns in the system inside, to pass through the obstructive environment. The system will operate under distributed communication topology to prevent scalability issues, which is relatively challenging as compared to a centralized system. There are three control objectives of this thesis, 1)., given a desired formation pattern set, find a suitable deformation strategy, so that the formation can adapt to the environments based on the given formation pattern set, 2)., maintain the rigidity of some desired formation through a distributed communication topology in real-time, 3)., guarantee collision-free paths for agents to fulfill reference trajectory tracking in the obstructive environment. Thus, we propose a novel leader-follower distributed nonlinear model predictive control (DNMPC) framework for realizing the above aims. In our framework, the leader is responsible for handling the first control objective by determining the level of deformation from a desired formation pattern to a so-called global reference information pattern, and the other two control objectives will be integrated into our DNMPC framework subject to an objective function and a corresponding constraint set. It is noteworthy that to minimize the rigidity cost in terms of an objective function while utilizing neighbors' information preserves the formation pattern in a cluttered environment, and viewing all the obstacles as hard constraints in the employed solver in our work guarantees collision-free paths of all agents during the course of the entire formation. In short, with the proposed formation framework based on DNMPC, the agents not only can achieve formation tracking control in the environment with obstacles but also are able to interact with the environment through the deformable formation scheme. Finally, we conduct several computer simulations and associated real-world experiments under various scenarios to demonstrate the feasibility and efficacy of our proposed approach.