博碩士論文查詢聯邦系統首頁 聯絡信箱 關於我們 加入我們

本論文已被瀏覽 58 次, [ 造訪詳細資料與全文 ] 13 次,[ 回到前頁查詢結果 ] [ 重新搜尋 ]

以類神經網路為基礎之教導機械人模仿人類動作的機制

A Neural-Network-based Mechanism for Teaching Robots to Imitate Human Actions

作者:林士傑
畢業學校:國立中央大學
出版單位:國立中央大學
核准日期:2013-08-30
類型:Electronic Thesis or Dissertation
權限:Copyright information available at source archive--National Central University....

中文摘要

個人化機器人被設計來在居家環境中能協助或娛樂人們,並且被期待
能夠與人們互動,因此,越來越吸引來自各領域的人們的注意力,而機器
人能夠模仿人類動作則被視為能與人類在動作上互動的第一步。本篇論文
提出一個以類神經網路為基礎的機制,得以讓機器人即時模仿人類動作。
主要由三個步驟組成:1)建立人體基本動作單元、2)針對每個人體基本動
作單元,利用最佳化演算法找出相對應的機械人關節的馬達角度、3)利用
前述步驟之結果,設計以類神經網路為基礎的機械人關節角度控制器。
首先由收集的多種人體動作序列,利用分群演算法找出基本動作單
元。由於人體基本動作單元數未知,我們採用非監督式的自我組織特徵映
射圖(SOM)方法,藉由人體動作資料的拓樸分佈,找出基本人體動作單元。
第二步是針對這些基本動作單元,找出對應的機械人關節馬達角度,使得
機器人的能作出與人體基本動作最相似的動作,此問題可被視為一個在高
維度空間中尋找一個最佳解(亦即最佳之馬達角度組合)的最佳化問題,會
隨著機器人之馬達數目的增加而大大增加其複雜度。針對此最佳化問題,
本論文特別發展了一個以鴿子覓食機制為構想的最佳化演算法─鴿子群體
最佳化演算法(Dove Swarm Optimization),希望透過此鴿子群體最佳化演
算法得以快速找到好的機械人馬達關節角度組合。第三步則是以上述找到
之機械人馬達關節角度對應之資料集作為訓練資料,訓練一個監督式的類
神經網路做為機械人關節角度控制器。本論文比較多層感知機網路以及放
射狀基底函數網路的結果,最後選用多層感知機網路作為控制器。
在實驗結果部份,本論文與線性對應(linear mapping)的方法比對差
異度、方均根誤差以及時間,結果發現整體效能平均提升約10%,而對於在
線性對應方式中表現特別不好的人體姿態,其改進比例則為13%,而在線性
對應方式中已表現很好的姿態則無太大之改進。

英文摘要

Personal service robots are designed to assist or entertain people in
domestic environments and expected to engage in social-human interactions;
therefore, they are gaining more and more attentions from many different fields.
A robot that can imitate human actions can be regarded as the first step for
interacting with humans from the viewpoint of actions. This dissertation presents
a neural-network-based imitation mechanism for teaching a robot to imitate
human actions. The proposed mechanism involves in the following three steps: 1)
the generation of basic human motion units, 2) the mapping between basic
human motion units and robot joint motor angles, and 3) the construction of a
NN-based controller.
First of all, we collected several human motion sequences consisted of
many different human activities and then used a clustering algorithm to cluster
the collected human actions into a set of basic human motion units. Since the
number of basic human motion units is unknown, we decided to adopt the
self-organized feature map (SOM) as the clustering tool to generate basic human
motion units. Secondly, for each basic human action unit, we need to find a
combination of robot joint motor angles to make the robot pose be similar to the
corresponding human pose. The problem can be regarded as an optimization
problem of which goal is to find an optimized solution (i.e., the best
combination of robot joint motor angles) in a multi-dimensional space. The
complexity of the optimization problem greatly increases as the number of robot
joint motors. To provide a good solution to the optimization problem, this
dissertation also proposes the new optimal algorithm called dove swarm
optimization (DSO), which is motivated by the doves foraging behavior. The
proposed DSO is adopted to affectively find the best combination of robot joint
motor angles corresponding to each basic human motion unit. In the third step,
the data set generated in the previous step is adopted as the training data set to
construct a NN-based controller. From our simulations, we found that controller
performance achieved by the multilayer perceptrons (MLP) outperformed the
radial basis function network (RBFN); therefore, we decided to adopt the MLP
to construct the NN-based controller.
The proposed mechanism was compared with the most straightforward
linear mapping method based on the root mean squared error and computational
time. In simulation results, we found that the proposed imitation mechanism
could promote the performance about 10% on average. The worst one hundred
basic human actions achieved by the linear mapping method, the imitation
performance could be improved to 13% by our mechanism. As for the best one
hundred basic human actions achieved by the linear mapping method, our
imitation mechanism did not clearly improve the imitation performance.


指導教授 - 蘇木春


 

計畫贊助者: