結合視覺與雷射建立拓樸式地圖應用於自主移動服務型機器人直覺式導航

Abstract

對於服務型機器人而言，移動能力是必備的功能之一。例如，我們可能要求服務型機器人協助我們遞送物品，或是從另一個房間拿取物品；在飯店，房客可以從房間訂餐，服務型機器人就必須有能力將餐點從廚房送至指定房間；在醫院，服務型機器人需要協助運送化學物質或是藥劑容器等等。我們可以想像未來公共場域中將有許多服務型機器人，那麼在這些情況下，服務型機器人就必須了解環境，並具備自動導航的能力。目前的導航方式多半倚賴數值地圖。一般而言，數值地圖是由SLAM(Simultaneous Localization and Mapping)方法所建立的。機器人可以利用數值地圖來規劃抵達目的地之最短路徑。然而，這些自動規劃出來的路徑並不一定是人類最能接受的路徑。有些路徑可能使機器人太靠近角落，或是穿越一些我們不想讓機器人進入的區域。換句話說，數值地圖僅紀錄了空間是否有障礙物等等資訊，卻沒有把人類喜好的路徑給一同紀錄下來。一旦我們希望將服務型機器人快速應用於某個環境時，這樣的缺點便會隨之彰顯。除此之外，數值地圖缺乏語意紀錄。一般而言，數值地圖包含了許多精確的座標。一旦我們希望機器人前往某個目的地，我們便需要給定這個目的地的座標點。這種情況就像當一個人想前往某個地方時，我們提供他經緯度。而機器人將會盡最大可能努力往這個座標點前進。因此，一旦有障礙物阻擋機器人抵達目標點，機器人便會持續認為他尚未抵達目的地，即使他早已抵達指定地點如”廚房”。為了能利用數值地圖以及拓樸式地圖的優點，我們提出一種混合式地圖。在這種混合式地圖當中，拓樸式地圖將負責紀錄語意，而數值地圖則是用來導航。我們將機器人所見影像儲存於拓樸式地圖中，這些資訊將用來進行機器人定位。而在我們進行建圖時機器人所行走的路徑，則會被紀錄為拓樸式地圖的edge。因此機器人在接下來導航時，就能夠規劃人類較能接受的路徑。我們使用類神經網路來比較機器人眼前影像以及儲存於拓樸式地圖中的影像，並且提出一個以影像為基礎的粒子濾波器，可以產生一種「語意姿態」，能夠讓機器人在定位上擁有更好的靈活度。拓樸式地圖以及「語意姿態」將能降低導航時間以及提昇導航成功率。我們在一800平方公尺的室內環境測試我們的演算法。我們紀錄了花費時間以及成功率，實驗結果顯示我們的成功率勝過傳統導航方式達16%，這樣的結果表示我們的方法將能讓機器人更穩定地進行導航。

The ability of navigation is a necessity for service robots.For example, we may ask a service to deliver objects, or take something in another room for us.In hotels, guests may order some meal in their room.Service robots need to be capable of carrying the meal from the kitchen to the room.Or in hospitals, service robots need to help deliver medicine or chemical containers.We can imagine that service robots are around us in every public area in the future.In these scenarios, service robots need to understand environments and navigate to destination safely and robustly. Navigation methods nowadays are mainly based on metric maps.Metric maps are normally generated by SLAM (Simultaneous Localization and Mapping) methods.With metric maps provided, robots can plan a shortest to destination with ease.However, these planned paths may not be the most human preferable ones.Some paths may be too close to corner, or passing through some unwanted areas.In other words, metric maps only record information of space occupation.Information of available and human preferable paths are not included.This can be a disadvantage once we want to make a service robot be fastly applied in any indoor environment.Moreover, metric maps are lack of semantic meaning.Metric maps are normally composed of precise coordinates.Once we want a service robot to do navigation, we also need to give it a set of coordinate.This situation is similar to providing longitude and latitude when someone wants to go to a place.In this case, semantic meaning is ignored and not suitable for intrinsic understanding.Without semantic meaning, a service robot may struggling in reaching a precise coordinate.For example, if we want a service robot to go to ``living room', we need to provide a set of coordinate when using metric maps.The robot will try its best reaching the goal coordinate.Therefore, once there is an obstacle stop the robot from reaching the goal coordinate, the robot will judge that it hasn't reached the goal even if it has already be in the ``living room'.Although metric maps are rich in details and suitable for navigation, it is hard to label abstract concept on it.For example, it is difficult to define an appropriate region on metric maps to represent a ``living room'.In the other hand, topological maps can store any data in their topological nodes, such as images or object labels.This can make it easier to integrate semantic meaning into maps.Nevertheless, topological maps are not precise enough.They cannot indicate precise spatial information compared metric maps.It is almost impossible to do navigation with only topological maps presented.To take advantage of both metric maps and topological maps, we propose a hybrid metirc-topomap.In this hybrid map, the topological map is responsible for semantic meaning recording, and the metric maps is used for collision avoidance.We store images in topological nodes, which is used for robot localization.The paths in mapping stage will also be recorded as topological edges.This help robots navigate in a more human preferable way in the future.We use a neural network to compare view from the robot and the images stored in the topological map, and then generate similarity values.We propose an image base particle filter to generate an ``semantic pose', which can give localization results more flexibility.With the help of topological maps and semantic pose, our proposed method can shorten navigation consumed time and make navigation with higher success rate.We test our algorithm in a 800 square meters indoor environment.We record the consumed time and success rate for the 5 paths in the environment.The experimental results show that the robot navigation success rate of our proposed method exceeds traditional navigation methods for about 16%.The result shows that our method can help service robots navigate more robustly.