機器人在導航時，在環境中的定位是困難但是卻基本必須的工作。本論文提出一個機器人定位以及地圖建構之系統，透過此演算法及系統，機器人能自動探索巡邏已學習之環境，或是規劃最短路徑來到達使用者事先指定的地點。 首先機器人要學習環境資訊前，會先進入探索模式。在探索模式中，機器人使用其視覺與紅外線感測器，建構一個包括節點與邊所表示的圖形地圖。節點代表一個特定地點（如岔路等），而節點間可到達之路徑則由邊所相連。在每個節點上，會將其環境影像作為訓練資料集。當探索結束回到原點時，這些訓練資料集會用來訓練一個多層感知機來記憶環境資訊。探索模式結束後會建立一環境地圖，接著會進入操作模式。 在操作模式中，機器人能自動巡邏環境且讓使用者能在遠端監控影像，或是執行使用者給予的特定地點巡邏。在導航過程中，機器人能夠依照之前所訓練的多層感知機，得知所在位置。最後，我們訓練SONY 所出產的愛寶（AIBO）機器狗來在居家環境中實際導航，來驗證我們的演算法與系統。 Robot localization has been a very challenging task in mobile robotics since it in essential for a broad range of mobile robot tasks. This thesis proposes a new vision-based robot localization and map-building algorithm. Via the proposed algorithm, a robot can automatically patrol the environment whose environment information has been learned or plan a shortest path to visit some particular locations pre-specifies by the user. To learn a new environment, a robot must first proceed to the exploration procedure (EP). In EP, a robot uses its vision and an infrared sensor to build a map of the unknown environment. The map is represented as a graph which consists of vertexes and edge. When a robot in navigating, a vertex is generated whenever a distinct environment (e.g. intersections, blind alleys, etc.) is detected and an edge is used to connect these vertex. At each vertex or particular location, images of the environment will be stared in a training data set. After the robot have finished the navigation tour and go back to the original starting position, a two-layer perceptron is trained to memorize the environment using the collected training data set. After the robot has built the environment map after the end of the EP, it enters the operation procedure (OP). In OP, the robot may automatically patrol the environment and transmit images to remote clients via a web browser or execute a particular patrolling task assigned by the user. During a navigation tour, the robot knows its location by contract a match between the observation and the expectation as derived from the database. The match is computed by feeding the observation to the trained MLP. Finally, the performance of the proposed algorithm is demonstrated by training a SONY AIBO to navigate a home environment.
