基於圖像嵌入與深度強化學習之無地圖室內視覺導航

Map-less Indoor Visual Navigation based on Image Embedding and Deep Reinforcement Learning

徐世曦(Shih-Hsi Hsu)

指導教授 ： 傅立成

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2017年)

https://doi.org/10.6342/NTU201704104

摘要

在傳統的視覺室內導航中，問題常常被敘述為如何從影像感測轉換成控制策略。而主要研究以分成兩種，一種方法會利用影像和預先建好的地圖來進行室內導航，此種方法可以達成全域的最佳規劃，然而地圖建置費時又需要大量的運算，而當環境變動時也不容易即時的更新；另一種方法則是嘗試只利用即時影像的方式來控制移動，可是此方法往往受限於資訊的局部性，而無法達成全域的最佳規劃。近年來深度學習相關的研究興起，許多領域受益於此而得以獲得長足的進步，而最近又以深度強化學習最受矚目。在室內導航問題中，深度強化學習可以幫助機器人將複雜的環境影像轉換成馬達控制的指令，藉以到達目的地。
本論文提出了一個新穎的架構，其目標在於達成無須地圖資訊也可以讓機器人學習如何在寬廣的室內空間中導航。寬廣空間中的室內導航往往牽涉到複雜的空間理解，尤其是室內空間往往有諸多的牆壁，門，遮蔽住視線，使得影像輸入往往複雜且難以分析。借助分散式深度強化學習與自動編碼器所產生的圖像嵌入空間，本方法可以實現機器人在寬廣的空間導航，不需藉助額外的地圖與人類的指導，可以到達指定的目的地。在實驗的部分中我們用模擬的室內環境和現實的室內環境檢驗了提出的方法，並且成功的達成了導航的任務。

關鍵字

深度強化學習 ； 圖像嵌入 ； 無地圖室內視覺導航

並列摘要

In the traditional vision based indoor navigation domain, the problem is usually described as how to convert the visual perception to control policy. The main research can be categorized into two types, one of which is the method using pre-constructed map and current perception to accomplish indoor visual navigation. This type of methods can easily achieve global optimal path planning, however, mapping algorithm is usually time-consuming and needs a lot of computation power. Also, this type of methods are sensitive to the changes in the environment and the map cannot be easily updated in a short time. Another types of method are navigation based on only real time image, yet this type of methods is restricted to local perception, and it is hard to achieve global optimal planning. Recently, as the rising of deep learning related researches, many research fields make a progress, and the deep reinforcement learning gets the most attention. In the indoor visual navigation problem, deep reinforcement learning can help the robot to convert the complicated environment scene to motor control command, and accomplish the navigation task.
In this thesis, we propose a novel structure, where objective is to achieve large-scale environment navigation in the indoor environment without pre-constructed map. The large-scale indoor environment needs good understanding to work for complex spatial perception, especially when the indoor space consists of many walls and doors which might occlude the view of robot. By the proposed distributed deep reinforcement learning and image embedding space generated by auto-encoder, out method can achieve large-scale indoor visual navigation without extra map information and human instruction. In the experiments, we show the validation of our proposed method by conducting successful navigation tasks both in simulation and real environments.

並列關鍵字

Deep reinforcement learning ； image embedding ； map-less indoor visual navigation

參考文獻

[2] Sim, Robert, and Gregory Dudek. “Learning generative models of scene features.” International Journal of Computer Vision 60.1 (2004): 45-61.

[3] Sim, Robert, and Gregory Dudek. "Effective exploration strategies for the construction of visual maps." Intelligent Robots and Systems, (2003).

[4] Sim, Robert, Pantelis Elinas, Matt Griffin, Alex Shyr and James J. Little. “Design and analysis of a framework for real-time vision-based SLAM using Rao-Blackwellised particle filters.” Computer and Robot Vision, (2006).

[5] Sim, Robert, and James J. Little. “Autonomous vision-based exploration and mapping using hybrid maps and Rao-Blackwellised particle filters.” Intelligent Robots and Systems, (2006).

[6] De Cristóforis, Pablo, Matias Nitsche, Tomáš Krajník, Taihú Pire, and Marta Mejail. “Hybrid vision-based navigation for mobile robots in mixed indoor/outdoor environments.” Pattern Recognition Letters 53 (2015): 118-128.