An application of scenario intelligence in dynamic environment exploration of office buildings

Ph.D.Scenario intelligence refers to the concept of constructing scenario-bounded artificial intelligence for performing in-scenario tasks. Scenario intelligence bases the development of intelligent systems on specific environments and focuses on improving their robustness, efficiency, and optimality.This thesis exploits the concept of scenario intelligence to tackle dynamic robotic exploration problem, in which a robot is autonomously navigated to build maps in office buildings. Four fundamental aspects of the dynamic exploration problem are discussed: map generation and motion planning in dynamic environments, decision making in partially known environments, and interaction with environment infrastructures.We first propose a distance transform algorithm for online map generation. The algorithm can incrementally integrate environment changes into the mapping process and outputs high quality map representations. The algorithm is implemented based on an efficient volumetric dataset and can effectively address large-scale environment mapping problem.We then present a novel motion planning framework for online trajectory generation in populated environments. The framework is comprised of a front end for searching optimal time assignments and a back end for generating smooth trajectories. We formulate the framework over a state-time space, based on which the obstacle dynamics are integrated into the planning process to help improve planning safety.Furthermore, we develop a decision-making framework for exploration goal selection based on the reinforcement learning approach. Compared with conventional methods, the proposed framework selects goals by inferring environment structures and thus can leverage more information about the environment to make better decisions.Lastly, we present an elevator button operation system to enable the inter-floor exploration of office buildings. The system contains a button recognition algorithm, a button pose estimation algorithm, and an operation controller. Through integrating these modules into a typical manipulation framework, the system is enabled to operate elevator buttons autonomously.The presented algorithms and frameworks constitute the fundamentals of dynamic robotic exploration of office buildings, and their implementations are typically scenario-dependent. By applying the concept of scenario intelligence, the robustness, efficiency, and optimality of these methods are significantly improved, though the generalization ability is sacrificed to some extent. Extensive experiments demonstrate the effectiveness of these methods in dealing with real-world applications and also verifies the philosophy conveyed by scenario intelligence.場景智能是指通過構造場景受限的人工智慧，以實現對行場景任務的執行。場景智能將智能系統的開發基於特定場景，著重提高系統的魯棒性、高效性和最優性。本文利用場景智能的概念來解決機器人在辦公樓中自主導航並繪製地圖的動態探索問題。討論了動態探索問題的四個基本方面：動態環境中的地圖生成和運動規劃、部分已知環境中的決策以及與環境基礎設施的交互。我們首先提出了一種用於線上地圖生成的距離變換算法。該演算法可以逐步地將環境變化集成到構圖過程中，輸出高質量的地圖表徵。該演算法基於高效的體數據集實現，能夠有效地解決大規模環境構圖問題。然後，我們提出了一個新的運動規劃框架，用於人員密集環境下的線上軌跡生成。該框架由一個用於搜索最佳時間分配的前端和用於生成平滑軌跡的後端組成。我們定義了一個基於狀態-時間空間的規劃問題，並在此基礎上將障礙物動力學融入規劃過程，以幫助提高規劃的安全性。我們進一步提出了一個基於強化學習方法的決策框架用於探索目標的選擇。與傳統方法相比，該框架通過推斷環境結構選擇目標，從而可以利用更多的環境資訊來做出更好的決策。最後，我們提出了一個電梯按鈕操作系統，以實現機器人在辦公樓的層間探索。該系統包括了按鈕識別算法、按鈕姿態估計算法和一個操作控制器。通過將這些模組集成到現有的機械臂操作框架中，我們實現了對辦公樓電梯按鈕的自主操作。本文中所提出的演算法和框架構成了辦公樓內動態機器人自主探測的基礎。這些算法和框架都是典型的場景依賴案例。通過應用場景智能的概念，雖然在一定程度上犧牲了泛化能力，但這些方法的魯棒性、高效性和最優性都得到了顯著提高。大量的實驗證明了這些方法在處理實際應用中的有效性，也驗證了場景智能所傳達的設計理念。Zhu, Delong.Thesis Ph.D. Chinese University of Hong Kong 2020.Includes bibliographical references (leaves 116-135).Abstracts also in Chinese.Title from PDF title page (viewed on March 21, 2022)