請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/79013
標題: | 互動式陪伴型機器人之人機信任提升系統 Human-Robot Trust Promotion System for Interactive Robot Companion |
作者: | Xingzhi Guo 郭行之 |
指導教授: | 傅立成 |
關鍵字: | 人-機器人信任,社交機器人伴侶,貝葉斯網絡,強化學習,常識知識圖譜, Human-Robot Trust,Social Robot Companion,Bayesian Network,Reinforcement Learning,Commonsense Knowledge Graph, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 作為『人-機器人互動』(人機互動)的關鍵要素,人對機器人的信任感能夠增 強人機互動的舒適度與可用度。本論文探討了一個能夠提升人對陪伴型機器人信 任感的人機互動系統。該系統能夠觀察人的臉部表情,當人的面部表情是負面的時, 其能觸發機器人主動給予人類關心,引發人的自我披露 (Self-disclosure) ,並基於 自我披露的內容產生一句適當的回覆。本論文著重於的問題有: (1) 機器人如何從 人的自我披露中推論出人的心理感受以便表達機器人的同理心; (2) 機器人如何產 生具有善意且符合互動個體之個人喜好的回覆。 該系統配備深度學習技術用以偵 測人的臉部負面表情,並以此作為訊號主動引出人的自我披露。以自我披露的內容 作為觀察到的資訊,本系統採用貝氏網絡 (Bayesian Network) 搭配建構的『具備因 果關係的常識知識圖譜』,並運用推論演算法 Gibbs Sampling Approximation Algorithm 來推論人的心理感受。結合自我披露的內容與推論出的心理感受,本系 統亦通過貝葉斯網絡推論出人可能會做的 (Desire) ,並以強化學習 (Reinforcement Learning) 的方式從互動歷史中掌握人的個人喜好,並以此為參考篩 選符合個人喜好的推論結果。基於自我披露的內容,推論出的人的心理感受 (Feeling) 與人想做的事 (Desire) ,本系統採用檢索式模型 (Retrieval-based model) 產生一個合適的回覆用於和人的互動,並進一步提升人對機器人的信任度。實驗結 果顯示本系統能夠正確推論出人的心理感受與想做的事,並且產生適宜的回覆內 容,最終通過互動提升人對機器人的信任感。 As a key component in Human-Robot Interaction (HRI), human’s trust toward robots can facilitate the HRI in terms of comfortability and usability. This thesis investigates a robot companion system which is able to promote human’s trust toward robots when human is in a bad mood by generating an appropriate response based on the human’s self-disclosure induced by the robot. This thesis focuses on 1). how the system infers human’s mental feelings implied from his/her self-disclosure in order to show empathy; 2). how the system generates a response for human while taking into account goodwill and personal preference. The system is equipped with deep learning techniques to detect human’s negative facial expression, which in turn can be used as a cue for robot to proactively induce human’s self-disclosure. Given the self-disclosure as observations, the system infers human’s feelings by applying Gibbs sampling approximation inference over a Bayesian Network which is constructed from commonsense knowledge. Together with the inferred feelings and the self-disclosure content, the system infers a set of human’s desires from the Bayesian Network and selects one of them based on the personal preference, of which the latter is learned from the past interactions through reinforcement learning. Based on the content of the self-disclosure, inferred feelings and desires, the system uses the retrieval-based model to generate an appropriate response to interact with human and ultimately promotes trust. The experiment results show that the proposed system can correctly infer human’s feelings as well as desires, and generate an appropriate response, resulting in the improvement of human’s trust toward the robot. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/79013 |
DOI: | 10.6342/NTU201803372 |
全文授權: | 有償授權 |
電子全文公開日期: | 2023-08-19 |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-R04922143-1.pdf 目前未授權公開取用 | 11.19 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。