본문

서브메뉴

Dexterous Robot Learning at Scale
Dexterous Robot Learning at Scale
Dexterous Robot Learning at Scale

상세정보

자료유형  
 학위논문 서양
최종처리일시  
20260202105056
ISBN  
9798288819056
DDC  
620
저자명  
Wang, Chen.
서명/저자  
Dexterous Robot Learning at Scale
발행사항  
[Sl] : Stanford University, 2025
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2025
형태사항  
127 p
주기사항  
Source: Dissertations Abstracts International, Volume: 87-02, Section: B.
주기사항  
Advisor: Li, Fei-Fei;Liu, Karen.
학위논문주기  
Thesis (Ph.D.)--Stanford University, 2025.
초록/해제  
요약Dexterous manipulation remains one of the most challenging and essential capabilities in robotics. It requires fine-grained, contact-rich control and generalization across diverse objects and environments. While recent advances in learning-based methods have demonstrated promise, progress is fundamentally limited by the availability of high-quality training data. In contrast to fields like vision or language that benefit from large-scale internet datasets, robot learning faces a "chicken-and-egg" problem: achieving robust robot manipulation necessitates large-scale, high-quality data, but the collection of such data typically requires already proficient robotic systems. How can we overcome this data bottleneck and unlock scalable training for dexterous manipulation? This thesis explores how to scale dexterous robot learning by leveraging diverse forms of human demonstrations as a rich, structured, and increasingly accessible data source. I begin with high-fidelity human teleoperation, which provides precise supervision but is costly and limited in scale. I then investigate passive third-person video as a widely available, low-cost alternative that offers broad task coverage despite lower fidelity. To bridge the gap between precision and scalability, I introduce a wearable motion capture system that enables detailed hand trajectory collection in unconstrained environments. To further address the embodiment gap between human and robot hands, I develop a hierarchical policy learning framework that leverages reinforcement learning in simulation to translate human wrist and finger motion into effective robot actions. Finally, I incorporate multimodal human signals-including muscle activity and audio-to enrich the learning signal for contact-rich and force-sensitive manipulation tasks. Throughout the thesis, I develop algorithms that distill key insights for control from human demonstrations, bridging the embodiment gap between human demonstrations and robotic execution. Together, these contributions offer a scalable framework for dexterous robot learning-one that brings together the richness of human behavior and the scalability of modern learning systems to advance general-purpose robotic manipulation.
일반주제명  
Robots
일반주제명  
Motion capture
일반주제명  
Success
일반주제명  
Computer science
일반주제명  
Computer engineering
일반주제명  
Robotics
키워드  
Dexterous manipulation
키워드  
Human teleoperation
키워드  
Multimodal human signals
키워드  
Robotic manipulation
기타저자  
Stanford University.
기본자료저록  
Dissertations Abstracts International. 87-02B.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008260126s2025        us                              c    eng  d
■001000017359293
■00520260202105056
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798288819056
■035    ▼a(MiAaPQ)AAI32201038
■035    ▼a(MiAaPQ)Stanfordzw936yz7062
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a620
■1001  ▼aWang,  Chen.
■24510▼aDexterous  Robot  Learning  at  Scale
■260    ▼a[Sl]▼bStanford  University▼c2025
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2025
■300    ▼a127  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  87-02,  Section:  B.
■500    ▼aAdvisor:  Li,  Fei-Fei;Liu,  Karen.
■5021  ▼aThesis  (Ph.D.)--Stanford  University,  2025.
■520    ▼aDexterous  manipulation  remains  one  of  the  most  challenging  and  essential  capabilities  in  robotics.  It  requires  fine-grained,  contact-rich  control  and  generalization  across  diverse  objects  and  environments.  While  recent  advances  in  learning-based  methods  have  demonstrated  promise,  progress  is  fundamentally  limited  by  the  availability  of  high-quality  training  data.  In  contrast  to  fields  like  vision  or  language  that  benefit  from  large-scale  internet  datasets,  robot  learning  faces  a  "chicken-and-egg"  problem:  achieving  robust  robot  manipulation  necessitates  large-scale,  high-quality  data,  but  the  collection  of  such  data  typically  requires  already  proficient  robotic  systems.  How  can  we  overcome  this  data  bottleneck  and  unlock  scalable  training  for  dexterous  manipulation?  This  thesis  explores  how  to  scale  dexterous  robot  learning  by  leveraging  diverse  forms  of  human  demonstrations  as  a  rich,  structured,  and  increasingly  accessible  data  source.  I  begin  with  high-fidelity  human  teleoperation,  which  provides  precise  supervision  but  is  costly  and  limited  in  scale.  I  then  investigate  passive  third-person  video  as  a  widely  available,  low-cost  alternative  that  offers  broad  task  coverage  despite  lower  fidelity.  To  bridge  the  gap  between  precision  and  scalability,  I  introduce  a  wearable  motion  capture  system  that  enables  detailed  hand  trajectory  collection  in  unconstrained  environments.  To  further  address  the  embodiment  gap  between  human  and  robot  hands,  I  develop  a  hierarchical  policy  learning  framework  that  leverages  reinforcement  learning  in  simulation  to  translate  human  wrist  and  finger  motion  into  effective  robot  actions.  Finally,  I  incorporate  multimodal  human  signals-including  muscle  activity  and  audio-to  enrich  the  learning  signal  for  contact-rich  and  force-sensitive  manipulation  tasks.  Throughout  the  thesis,  I  develop  algorithms  that  distill  key  insights  for  control  from  human  demonstrations,  bridging  the  embodiment  gap  between  human  demonstrations  and  robotic  execution.  Together,  these  contributions  offer  a  scalable  framework  for  dexterous  robot  learning-one  that  brings  together  the  richness  of  human  behavior  and  the  scalability  of  modern  learning  systems  to  advance  general-purpose  robotic  manipulation.
■590    ▼aSchool  code:  0212.
■650  4▼aRobots
■650  4▼aMotion  capture
■650  4▼aSuccess
■650  4▼aComputer  science
■650  4▼aComputer  engineering
■650  4▼aRobotics
■653    ▼aDexterous  manipulation
■653    ▼aHuman  teleoperation
■653    ▼aMultimodal  human  signals
■653    ▼aRobotic  manipulation
■690    ▼a0771
■690    ▼a0984
■690    ▼a0464
■71020▼aStanford  University.
■7730  ▼tDissertations  Abstracts  International▼g87-02B.
■790    ▼a0212
■791    ▼aPh.D.
■792    ▼a2025
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359293▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

미리보기

내보내기

chatGPT토론

Ai 추천 관련 도서


    신착도서 더보기
    최근 3년간 통계입니다.

    소장정보

    • 예약
    • 소재불명신고
    • 나의폴더
    • 우선정리요청
    • 비도서대출신청
    • 야간 도서대출신청
    소장자료
    등록번호 청구기호 소장처 대출가능여부 대출정보
    TF15245 전자도서 대출가능 마이폴더 부재도서신고 비도서대출신청 야간 도서대출신청

    * 대출중인 자료에 한하여 예약이 가능합니다. 예약을 원하시면 예약버튼을 클릭하십시오.

    해당 도서를 다른 이용자가 함께 대출한 도서

    관련 인기도서

    로그인 후 이용 가능합니다.