全自動機器人具同時定位與建圖及綁架回復機制之研究

Abstract

在人機互動研究領域中，機器人被綁架問題及自動回復並到達目標是一個非常有趣且重要的研究議題。有別於工業用機器人，在行進中所工作環境通常是靜態的或是可被預測的。服務型機器人的工作環境就是我們的生活環境，在這個環境中存在著許多不確定性是機器人必須要克服的。機器人綁架問題就是眾多不確定性中其中一項。具體來說，因為機器人就在我們生活周遭，有的時候會我們會搬動機器人到其他地方為我們工作，清潔機器人就是最好的例子。當我們搬動機器人到其他地方的時候，對機器人是相當困擾的，因為它完全不知道發生什麼事，以及如何應對。 這篇論文主旨在於解決機器人綁架問題。我們提供一個演算法讓機器人知道其被綁架了，進而了解被綁架到何方使用一些機器人定位的方法。本論文所使用之主要的感測器為雷射測距儀。到目前為止蒙地卡羅定位法(Monte Carlo Localization)，為一個廣泛被使用的機器人定位方法，然而蒙地卡羅定位法在全域定位(Global Localization)的計算量非常龐大，耗時也相當長。一般而言全域定位功能在機器人綁架問題中扮演重要腳色。我們在這篇論文提出一個方法提升全域定位的計算效率同時降低計算時間。此方法結合了傳統蒙地卡羅定位法和現代機器學習方法 Fast Library for Approximated Nearest Neighbors (FLANN). 由於FLANN 需要定義描述子(descriptor)，我們在這篇論文也定義了一個新的基於雷射測距儀的資料格式描述子。我們將此描述子命名為Geometric Structure Feature Histogram (GSFH)。透過結合現代機器學習法FLANN與傳統蒙地卡羅定位法，我們大大降低了蒙地卡羅的計算負擔，並與提高計算效率。讓機器人綁架問題得以解決。實驗結果以及模擬結果都顯示了我們的方法的有效解決機器人綁架及自動回復到達原目標問題。

The Kidnapped Robot Problem is one of the essential and interesting issues in Human Robot Interaction research fields. Unlike industrial robot works in factory which is mostly in a static and predictable environment. Service robot works in the environment together with us, and there are many unpredictable factors the robot should overcome. Kidnapped Robot Problem is one of the unpredictable cases. Since the robot works around us, sometimes we may take up the robot to the other place in order to help us to deal a desirable task, for example cleaning robot. However, these actions may cause big problems because the robot does not have any idea what has happened. This thesis addresses the problem of the position and orientation (pose) recovery after the robot being kidnapped, based on Laser Range Finder (LRF) sensor. By now the Monte Carlo Localization (MCL) has been introduced as a useful localization method. However the computational load of MCL is extremely large and not efficient at the initial few steps (global localization), which causes the localization process to take long computation time after the robot has been kidnapped and resets the particles. This paper provides a methodology to solve it by fusing MCL with Fast Library for Approximate Nearest Neighbors (FLANN) machine learning technique. We design a feature for LRF data called Geometric Structure Feature Histogram (GSFH).The feature GSFH encodes the LRF data to use it as the descriptor in FLANN. By building the database previously and FLANN searching technique, we filter out the most impossible area and reduce the computation load of MCL. Both in simulation and real autonomous mobile robot experiments show the effectiveness of our method.