本論文提出一套即時閃避靜態與動態障礙物之安全避碰路徑規劃演算法，此高效率路徑重新規劃演算法應用於協作型機械臂在人機共工環境中執行避碰規劃。在機械臂執行既定任務初始路徑同時追蹤環境中的靜態與動態障礙物，當動態障礙物進入機械臂的工作空間時，本演算法將重新規劃一條安全避碰之局部路徑使機械臂避開靜態與動態障礙物。為了實現快速的局部路徑重新規劃，本論文提出一套混合人工勢場與快速隨機搜索樹之路徑規劃演算法，並同時保留此兩種演算法的優點。

本論文提出一套將CT影像與患部進行配準之方法，讓擴增實境投影虛擬模型到患部，並讓手術機器人進行準確的導航操作。為了提高表面模型配準的精度，本論文依據配準時兩點雲間疊合部分的對應關係，提出一個新的迭代最近點演算法(iterative closest point, ICP)，在原來的演算法中加入互相對應距離限制式以提升配準精度，實驗結果顯示所提出之演算法平均比Symmetric-ICP演算法減少68.5%配準誤差。本論文將所提出之配準方法整合擴增實境系統應用於手術機器人，並提出一套擴增實境與外部導航校正方法，以透過導航系統追蹤物件即時更新虛擬物件投影位置，以提供醫生更佳的手術輔助。

本論文提出一套移動式機器人基於視覺同時定位與地圖建立(visual SLAM)之改良設計，在傳統 ORB-SLAM2演算法之追蹤執行緒之追蹤執行緒 (Tracking)中嵌入動態人員移除之演算法，以提高自主移動機器人，以提高自主移動機器人，以提高自主移動機器人，以提高自主移動機器人 (Autonomous Mobile Robot, AMR)在有人員走動環境中在有人員走動環境中在有人員走動環境中之定位準確度。本論文以人員實例分割 (Instance segmentation)與運動追蹤(Motion tracking)之資訊，辨別影像中人員之運動狀態，將屬於動態人員分割區域中之ORB特徵點移除，並將此演算法整合到ORB-SLAM2之執行緒中。

現實世界的機器人經常面對雜亂堆疊的夾取情境，使夾取成功率隨環境複雜度的增加而大幅下降。為了有效地完成新穎物件抓取的任務，本研究提出了一種新的物件感知策略，它考慮包含目標物件以及環境資訊的盒狀點雲，透過遮罩分割作為其輸入的部分特徵。本論文提出一套提出之平行夾爪可用性(Affordance)感知模型能夠學習到目標物件與周圍環境的相互關係，用來濾除位於目標物件上低成功機率的夾取區域。具有場景理解的可用性(Affordance)模型，可以為物件層級的夾取系統帶來更合適的夾取區域抉擇，提高夾取之成功率。

近年來自主移動式機器人(Autonomous Mobile Robot, AMR)已開始被導入各種智慧服務與智慧製造場域中。未來AMR將更進一步與人員協作執行各種任務，達到更彈性與智慧之機器人自動化。本計畫之主要目標在開發一套協作型AMR之人機互動路徑規劃系統，透過即時感測之3D點雲模型，提供協作型AMR之機械臂與移動平台全方位路徑規劃，並以優化避碰路徑、縮短路徑偏移量及即時執行效率等性能指標，提升協作型機器人之使用彈性與實用性。本計畫團隊與達明機器人股份有限公司(Techman Robot Inc.)合作，未來將進行實際應用系統整合與測試，預計將所開發之模組技術轉移至合作廠商，實際應用至產業端。

本論文提出一套基於臉部辨識與頭部姿態估測之人機互動規劃， 使移動式機器人能對環境中特定人員拍攝照片。使用 RGB網路攝影機，偵測出畫面中的臉部資訊，畫面中的臉部資訊再以 深度學習類神經網路（ 深度學習類神經網路（ 深度學習類神經網路（ Deep Neural Network, DNN） 模型進行人臉辨識，進而計算欲拍照的特定對象之臉部姿態。論文中以機器人欲拍照目標對象 之位置與頭部姿態為依據，結合 移動式機器人自主定位功能，計算出目標人員之位置與朝向，位置與朝向， 規劃機器人拍照位姿及移動路徑 以達成拍攝照片之功能。透過所設計之人機介面，使用者以手機應用程式設定 機器人欲拍攝之目標對象，自主移動至 適當之拍照位置進行攝影。

本論文提出一個多機器人導航系統設計，使每一台機器人能在複雜且未知的環境中，不會彼此碰撞且有效率的到達目標點。為了能即時獲得每一台機器人的資訊並且有效的協調各機器人的行為，提出使用雲端運算實作多機器人避碰路徑規劃。導航架構分為三個部分，分別為多機器人避碰路徑規劃、障礙物閃避以及導航控制。其中多機器人避碰路徑規劃系統是透過雲端運算以A*演算法先對每一台機器人規劃到目標點之路徑，再結合Extended ORCA演算法進行優化與協調，為每一台機器人產生一條不會碰撞環境中靜態物體及其它機器人的最佳路徑。Extended ORCA演算法則是基於ORCA演算法延伸設計，本論文將速度資訊轉換為位置資訊，以產生多機器人之無碰撞路徑。導航系統，能讓每一台機器人在未知之複雜環境中，自主規劃其無碰撞路徑，安全且有效的完成導航任務。經由多台機器人實驗結果驗證了無論環境中是否具有不預期障礙物，每一台機器人都能透過本論文所提出的方法，毫無碰撞的到達所指定的目標點。

本論文提出一套語意同時定位與建圖(Semantic Simultaneous Localization and Mapping, Semantic SLAM)演算法，透過結合物體姿態估測與視覺同時定位與建圖(visual SLAM, vSLAM)演算法建立一個多層次語意地圖(Multi-layered Semantic Map)，並應用於機器人路徑規劃與自主導航。物體姿態估測演算法採用一套深度學習模型DenseFusion，利用RGB-D攝影機進行即時性的物體姿態辨識與估測，而vSLAM採用RTAB-Map優化產生3D稠密型地圖。為了實現自主導航控制，本論文利用點雲地圖壓縮方法得到一個多層次2D語意地圖，以供機器人路徑規劃與定位導航使用。機器人使用LiDAR和RGB-D攝影機進行機器人定位，前者藉由自適應性蒙地卡羅演算法(AMCL)與2D地圖匹配來獲得自身位置，後者則使用上述之物體姿態辨識演算法來得到機器人與環境物體的相對位置。藉由語意地圖中環境物體資訊，機器人可以規劃出一條適當的移動路徑，搭配以上定位資訊導航至目標位置。

本論文提出一套基於移動式機械臂(Mobile Manipulator)之混合相機系統(Hybrid camera system)手眼校正(Hand-eye calibration)設計，使一移動式機械臂能夠在靠站後，快速並同時完成Eye-in-hand及Eye-to-hand兩種相機系統的手眼校正。本論文對eye-in-hand/eye-to-hand混合相機系統結合李代數近似(Lie Algebra Approximation)進行手眼轉換旋轉部分之疊代，接著利用疊代出之精準解配合混合相機系統校正模型，針對手眼轉換的位移向量進行求解達成手眼校正。本論文推導出一套基於四元數轉換之旋轉初始值估測方法，針對混合相機系統校正模型 AXB=YC 進行初始校正，以求得手眼轉換旋轉部分的估測解，確保後續疊代過程中的收斂性質。為了減少手眼校正流程中資料收集的時間，本論文設計一3D校正標籤裝置，透過多個視覺標籤的重組來使相機能夠一次擷取多筆資料，以達到加速資料擷取的效果。本論文以六軸協作型機械臂TM5M搭配自行研製之移動平台進行實驗驗證。

本論文提出一種基於技能評估之分享控制器設計，以協助外科醫生藉由機器人輔助進行內視鏡手術。在機器人輔助腫瘤切除手術過程中，醫生為了保持切緣(resection margin)一致，需要執行精確的動作，而手術技能的差異會影響微創手術(Minimally Invasive Surgery, MIS)的品質。本論文結合技能評估與分享控制，使機器人自動化輔助MIS 操作過程，以期能夠執行更好的手術。我們透過訓練卷積神經網絡(Convolutional Neural Network, CNN)建立線上(on line)技能評估模型，並將訓練結果應用於調整遠端控制以及機器人自動化之間的分享控制權重，達成即時的動作輔助。本論文針對輪廓追蹤任務的分享控制器進行設計，模擬腫瘤切除手術動作。透過實驗室的手術機器人實驗平台，論文中以實驗驗證所提出方法的有效性。經由多個輪廓追蹤實驗結果顯示，與單純遠端操控相比，本論文提出的分享控制器所執行的輪廓追蹤可以提升追蹤精度34.5 %。

本論文提出一套基於深度神經網路 Deep Neural Networks , DNN 物體辨識之視覺定位與建圖 Visual Simultaneous Localization and Mapping , vSLAM 演算法將其應用於機器人路徑規劃與控制。在掃地機器人的應用我們需要使用一個耗電少和體積小的嵌入式系統做為運算平台因此，利用 RGB D 攝影機偵測環境，並開發輕量化ORB SLAM2 的系統 架構 使其能在 CPU 的運算平台上執行。透過加Streaming SIMD Extensions , 技術 達成 減少特徵點擷取演算法的運算時間，以及融合DNN物體辨識技術減少特徵點匹配之運算時間 ，達成執行於嵌入式系統上的目標 。

本論文針對協作型機械臂提出終端持有工具時之即時安全避碰設計。所提出的避碰設計分為三個部分，分別是動態物體追蹤、終端工具點雲追蹤與預測以及即時避碰運動規劃。本論文以KinectV2深度攝影機對環境中物體與人員進行追蹤，接著透過K-D tree演算法分割工具點雲與障礙物點雲。針對終端工具的點雲，我們提出一套演算法對終端工具建立模型並預測其下一時刻之姿態，解決攝影機視線遮蔽的問題。在即時安全避碰設計方面，本論文提出一個新的人工力場演算法(Artificial Potential Field, APF)使機械臂達成更靈活地閃避。本論文將終端工具整個體積納入考量，除了在障礙物與終端工具間產生排斥力(Repulsive force)及在目標點與終端產生吸引力(Attractive force) 以外，我們提出旋轉扭矩(Virtual torque)的設計並將其加入人工力場，使機械臂之終端工具得以旋轉來調整姿態，提前預防以降低發生碰撞的機率，也可以使為了閃避而產生的額外動作變得較為平順，縮短避碰路徑。

本論文提出一套基於擴增實境(Augmented Reality, AR)之機器人輔助手術導航系統。結合機器人準確穩定的特性及AR可視化輔助之監控，本系統可以在手術前根據3D醫學模型規劃機械臂的軌跡，並讓外科醫生可以透過擴增實境監督手術過程中的執行效果。為了實現圖像引導手術，本論文開發出使用基於標記之手眼校正，得出空間座標系與機械臂座標系間之轉換，達成手術機器人影像導航控制。本論文以機械臂鑽骨手術任務為設計範例，提出一個基於AR可視化及3D醫療模型之即時監視與調控設計，以協助醫師於術中監督及調整機械臂之執行軌跡。本論文以自行設計之AR標記，進行智慧眼鏡虛擬物件投影，並提出一種座標轉換演算法，達成AR虛擬座標系與世界座標系對位匹配，將監控UI介面虛擬軌跡規劃轉換成機械臂之軌跡。

本研究 發展一套基於擴增實境之(Augmented Reality, AR)內視鏡扶持機器人操控系統，讓醫生可以直接使用頭部轉動操控內視鏡影像畫面，同時能更加直覺進行手術動作。本論文設計 出一套頭控系統， 使用AR眼鏡的陀螺儀測量頭部轉動來控制內視鏡視野移動，並且藉由AR眼鏡同時顯示內視鏡影像以及現實畫面， 讓醫生在操作過程中不但能持續 直視內 視鏡的畫面，也能完整掌控手術房的實際 情形。由於頭控方法無法直覺控制內視鏡的 深度，本論文設計使用器械辨識的方法實現內視鏡深度控制 。並且為了讓內視鏡自動的深度調整不干擾到手術的進行，將先行判斷醫生是否在進行手術 動作 ，如果判斷醫生並非在進行手術的操作而是在觀察患部時，才可以進行內視鏡的深度控制。

本論文提出一套基於視覺之手術器械姿態辨識方法，期望能針對醫生進行微創手術 訓練時，單純利用視覺將手術器械之姿態估算出來，另一方面透過示範動作之學習未 來可應用於與標準手術動作之手術器械姿態比對，進行訓練效果評量。使用機器人進 行手術動作重現。本論文使用自行安裝之立體攝影機，配合在手術器械關節點上之顏 色標記，從影像中擷取手術動作中的器械關節點位置，搭配立體攝影機所產生的深度 資訊疊合出器械關節點的三維位置，再透過卡爾曼濾波器(Kalman Filter) 後得到的器械 關節點座標。本論文提出運動學演算法推算出手術器械四個關節轉動角度及機械臂六 自由度之位置座標儲存串列之座標點位置做為手術動作之資料。

本本論文主要針對醫療機器人領域進行研究。論文中提出以硬體加速設計提升逆向運動學計算，使能在精度範圍內達到最快之控制速度，改善外科手術機械臂之控制架構。硬體加速器以FPGA電路實現逆向運動學計算，透過優化設計達成縮小FPGA中邏輯電路模組並達成提高6自由度機械臂逆向運動學計算精度的整合設計，能在給定機械臂終端之卡氏座標回推到六軸馬達對應之角度。本論文使用National Instruments myRIO作為系統整合架構的基礎，其中包含在CPU上運行軟體之驅動程式及FPGA電路之硬體加速器。本研究透過點到點、線性以及圓形軌跡等實驗對提出之逆向運動學進行驗證。論文中將16位元、24位元和32位元之FPGA版進行精度之比較，結果顯示在32位元中，其最大誤差為0.2mm，而平均誤差為0.07mm。速度方面16位元僅需11微秒，24位元為28微秒，32位元也只需要35微秒之處理速度。

本計畫透過3D視覺，發展一套協作型機械臂即時避碰之工件取放控制模組，達成安全抓取功能，架構圖如左圖所示。機器人3D視覺主要分為兩部分，首先，使用RGB-D深度攝影機透過深度學習(Deep Neural Network, DNN)進行工件辨識，當工件在RGB影像中被成功辨識後，將點雲資料和Bounding box的辨識資訊疊合，分割出工件的點雲資料，進而估測工件姿態，進行機械臂bin picking。在協作機器人安全部分，機械臂執行工件抓取任務的過程中，其工作區以深度攝影機3D視覺監控，當有人員或障礙物進入工作區時，機械臂進行即時障礙物閃避，達到安全抓取的功能。 本計畫於達明機器人公司之TM5上進行整合測試，驗證系統的有效性。機械臂進行抓取任務時具障礙物閃避之功能，當人員或物品離開危險區域後，TM5會回到原本的工作路徑上繼續完成工件抓取任務。

This thesis presents a deep learning based object recognition and point cloud segmentation design for bin-picking. The deep learning design for point cloud segmentation is used to increase the efficiency of object recognition and reduce the computational efforts of pose estimation. A RealSense SR300 RGB-D camera is utilized to acquire the RGB images and depth images. RGB images are used by Deep Neural Network (DNN) object detector and the depth images are applied to generate point cloud of workpieces. When the object is recognized in the RGB images, it will be located by a bounding box. We propose a 3D-To-2D projection and index filtering scheme to segment the corresponding point cloud according to the bounding box. Then, a voting-scheme is employed to the segmented object point cloud to obtain coarse pose of the object. Then the pose clustering and iterative closest point (ICP) algorithms are used to refined the pose estimation. The estimated object pose data are sent to the robot controller for picking the object. Practical experiments on TM5 robot arm show that the average pose estimation time is less than 1.2 seconds, the average picking cycle time is less than 15 seconds and the picking success rate achieves 92.5%.

This thesis presents a novel indoor localization design based on bluetooth low energy ( wireless signal devices. With the location information, a user can call the robot to the user's location to provide service through a portable device such as smart phone . The accuracy of indoor localization of a specific person is one of the key elements of the system. In this thesis, deep neural network (DNN) is used to classify and recognize spatial regions based on receive signal strength indicator (RSSI). In order to improve the performance of position estimation and tracking, a particle filter is employed to increase accuracy of the wireless signal indoor localization system. The experimental results show that the proposed wireless signal indoor localization system can effectively track the position of people in the environment with accuracy of 50cm.

實驗室研發一套智慧醫院接待機器人，主要功能包含醫院大廳的接待、協助病患就診與健檢導引。現在一般大型醫院場域較廣，對病患及家屬容易造成不便及就診時之焦慮，藉由接待機器人之指引與導引將有助於人們了解醫院進而增加親切感。另一方面，健康保養與疾病預防觀念日益普及，健檢已是醫院重要之服務項目之一。大型醫院皆有健檢的服務，急需醫護人員提供協助，若有機器人提供健檢導引，將是邁向智慧醫院一項重要指標。 實驗室針對此問題進行改善，研發一台與網路結合的醫院機器人，結合雲端計算、排程規劃和機器人導引控制，有效的分配檢查的時間和減少需要醫護人員的勞力，提升醫院的健檢的品質和效率。機器人具有定位導航、障礙物閃避、排程規劃、人員偵測等功能模組。當機器人接收到來賓註冊健檢後，對所有使用者進行排程與路徑規劃，依序導引至健檢站，並且配合使用者移動速度進行導引，依照排程執行任務直到完成健檢項目。

This thesis presents a real-time motion planning and control design of a robot arm for human-robot collaborative safety. Multiple KinectV2 depth camera are utilized to model and track dynamic obstacles (e.g. Humans and objects) inside the robot workspace. Depth images are applied to generate the point cloud of segmented objects in the environment. A K-nearest neighbor (KNN) searching algorithm is used to cluster and find the closest point from the obstacle point cloud to the robot. Then, a Kalman filter is applied to estimate the obstacle position and velocity. For the collision avoidance behavior, the repulsive force is generated for the robot end effector by calculating the minimum distance between the robot and moving obstacles. Moreover, a novel collision-free motion planning method is proposed not only to keep robot body from colliding with objects but also to preserve the execution of robot’s original task under the Cartesian constraint of the environment. Practical experiments show that the 6-DOF robot arm can effectively avoid obstacles in a constrained environment and complete the original task.

Human voice provides a convenient and natural way to communicate with a service robot. It is desirable to operate a mobile robot using voice commands. However, in practice, simple voice commands are not sufficient to give continuous and safe motion of a mobile robot in a complex environment. This work studies a shared-control approach to remote control of a mobile robot. A remote control scheme for an omnidirectional mobile robot is proposed to overcome the delay-time problem of voice commands and maintain the safe motion of the robot using simple voice commands. In this work, we combine the robot autonomy and human intention by using shared-control techniques. The human and robot interactive gains are obtained by computing posterior of free-space using a Bayesian recursive algorithm. Experimental results are presented to demonstrate the effectiveness of the proposed method.

The objective of this work is to design and implement attitude control for an octocopter. Octocopter is one of the Unmanned Aerial Vehicles (UAVs), which have grown fast recently, and various applications have been developed. In this work, the dynamic model of an octocopter is obtained by making the configuration similar to a quadcopter. In order to stabilize the attitude control of the octocopter, we proposed a Fuzzy+PID design of the attitude controller. Both computer simulation and real-time experiments have been carried out by using LabVIEW programming. In the simulation, we investigated the performance of the proposed Fuzzy+PID controller and tuned the controller parameters. For real-time control, NI-MyRio has been chosen to implement the controller onboard the octocopter. The practical experiments show that the proposed Fuzzy+PID controllers can handle the disturbances during flying and have robustness against certain environmental conditions. The experimental results validate the simulations and verify the effectiveness of the developed controller.

This thesis presents a new method to provide compliant motion of a walk assistant robot based on user status and motion intent. A laser scanner is used to detect the user's gait information and a force/torque sensor detects desired motion direction. The compliance controls will provide the walking assistance. The robot velocity is regulated by user gait information and the robot motion direction is adjusted by force/torque sensor information. These two compliant controllers allow the walking assist robot to generate passive behavior and provide walking assistance effectively.

It is usually difficult to remotely control of a mobile robot using only an onboard camera. The robot can easily bump into an obstacle due to limited view angle of the on-site scene. The thesis presents a shared-control architecture for effective remote control an omnidirectional mobile platform. The platform is able to move in any direction without changing robot’s heading, such that a stable camera scene is available for remote control. A user interface is designed to use smart phone/tablet to remote control the robot by the interactive touch screen. The proposed controller determines the human and robot interactive gains by computing user’s confidence factor.

In this thesis, we propose a CAD-based 6-DOF pose estimation design for random bin-picking of multiple objects using a Kinect RGB-D sensor. A voting-scheme was adopted for 6-DOF pose estimation as well as recognition of a set of 6-DOF poses of different types of objects in the bin. We combine 3D CAD model of objects with a virtual camera to generate point cloud database for pose estimation. A design based on voxel grid filter is suggested to decrease the number of 3D point cloud of object for reducing time of pose estimation. Furthermore, we use an outlier filter to filter out bad matching poses and occluded ones.

In this thesis, a system is designed and experimented for efficient mobile manipulation of a dual-arm robot. Because of limited workspace of the robot arm, robot grasping may fail if the target object is outside the workspace. At the same time, navigation errors also affect the grasping as the robot needs to move toward the target object. In this thesis, we construct a graspability map to describe the possibility of successful grasping inside workspace of the robot arm and we also construct a reachability map to describe the feasibility the robot can navigate to the target. These two maps are integrated to determine the robot location for grasping the target object.

This thesis aims to develop a method to estimate contact force for minimally invasive surgery (MIS) by using image processing and 3D reconstruction modeling. In order to provide surgeons a feeling of contact force between human tissue and instrument, we develop a method to calculate tissue deformation by reconstructing 3D model and estimate force information. The force information will be feed-back to a haptic device to allow a surgeon feel contact force. In this thesis, we propose a 3D model reconstruction method to calculate tissue surface under deformation, and use the deformation data to obtain contact force near the instrument tip.

A location aware system provides location information of objects, users and mobile robots from sensor nodes deployed in the intelligent environment. The information can be used to support various intelligent behaviors of a service robot in day-to-day application scenarios. This thesis presents a probability-based approach to building a location aware system. With this approach, the inconsistencies often encountered in received signal strength indicator (RSSI) measurements are handled with a minimum calibration.

The objective of this study is to design an image tracking algorithm for the endoscopic system in Minimally Invasive Surgery (MIS). The endoscopic robot autonomously adjusts its pose according to the position of the instruments in image plane, and moves the endoscope to provide a suitable field of view. A method is proposed to identify the tip of instruments without using extra artificial markers.

This paper presents a mobile manipulation and visual servoing design for a configurable mobile manipulator by using a Kinect sensor. The goal of this method and design is to execute mobile manipulation tasks in dynamic environments. To achieve this goal, a mobile manipulation and visual grasping design is combined a vSLAM (visual simultaneous localization and mapping). The vSLAM method is proposed based on extended Kalman filter (EKF) and a Kinect RGBD sensor.

The goal of this study is to develop a robot (face) that can present proper emotional expression to a user in response to his/her emotional state.

The development of a robotic interactive emotional brain (IEB) can handle emotional state generation of a robot.IEB allows the robot to autonomously generate intimate and natural emotional behaviors in response to user’s emotional state. Different emotional characteristics of a robot can be designed-in an IEB.

The goal of this study is to develop a walk-assistive system for the elderly people. The robot should comply with the user’s motion intent without using a force/torque sensor.

The gravity effect is especially considered and compensated by the force observer such that the robot posses compliance on an slope surface.