擬人之行走：依第一人稱視角進行社交導航方式之全向輪機器人

Abstract

本文針對全向性移動機器人平台，提出了一種基於深度學習的社交友善型的室內導航系統。目的有二，第一是直接利用機器人在第一人稱視角感知的圖像中產生與人互動的行走策略，以便實現從人類行走方式學習到的社交友善型運動方式；第二是要结合行人躲避算法和静态障碍物躲避算法以完成导航任务。為了實現這一目標，首先會佩戴RGB相機和IMU，分別收集行人的骨骼資訊以及佩戴者的行走策略。之後，通過預處理收集到的資料來構建行人动态和避障行为資料集。然後利用行人骨骼資訊和佩戴者運動資訊作為輸入來訓練長短期記憶網路(LSTM)以獲得機器人運動策略。此外，結合基於鐳射的同時定位與建圖(Laser-based SLAM)，通過構建導航决策模擬模型，來實現複雜室內環境中的靜態障礙物躲避和行人躲避。 此方法克服了先前方法的缺點，如基於傳統導航技術的方法，因為它們只關注人的位置和社交力關係，體現了較弱的社交導航能力並且可能會引起互動人的心理不適，造成一種被侵犯的感覺。另外，為了使預測結果更加準確和穩定，我們創新性地設計了可以描述人類行走意圖的特征，並且收集了第一個使用監督學習預測此類問題的數據集。實驗表明，本文提出的方法可以顯著提高多人環境中的社交接受程度和社交友好程度。此外，與之前的研究相比，通過本文提出的端到端的多流度長短期記憶網路(Multi-stream LSTM)模型，可以獲得更好的實驗結果，提高預測準確度。

This thesis proposes a learning-based social navigation method in populated indoor environment for an omnidirectional robot. The aim is not only to produce an avoidance strategy directly from the images the robot perceived in first person view in order to perform the social friendly motions which are learnt from human, but also to combine the human avoidance and static obstacle avoidance in order to finish the whole navigation objective. To achieve this goal, we first wear the RGB camera and IMU to collect information of pedestrian skeleton and the walking strategy of collector, respectively. Henceforth, we build a human traveling dataset via preprocessing the data we collect. Then we utilize the human skeleton and distance as inputs to train an LSTM model to obtain robot motions. Furthermore, through the laser-based simultaneously localization and mapping to realize static obstacle avoidance and dynamic human avoidance simultaneously in complicated indoor environment. We overcome the disadvantages of previous methods like navigation techniques-based approaches, as they only focus on the human position and social force relation, which perform weak social navigation ability and may offend human psychological comfort. In addition, in order to get a better and stable performance we innovatively design specific features to describe the intension of human. We also collect a dataset which is used to solve this kind of supervised learning problem for the first time. The experiments show that our method can greatly improve the level of social acceptance and the level of social friendliness in populated environment. Besides, we can get a better results through our proposed end-to-end multi-stream LSTM model as compared with the previous models.