서브메뉴
검색
Data-Centric Generalization for Robot Manipulation: Simulation, Augmentation, and Adaptation
Data-Centric Generalization for Robot Manipulation: Simulation, Augmentation, and Adaptation
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103610
- ISBN
- 9798288865503
- DDC
- 629.8
- 서명/저자
- Data-Centric Generalization for Robot Manipulation: Simulation, Augmentation, and Adaptation
- 발행사항
- [Sl] : University of California, Berkeley, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 243 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
- 주기사항
- Advisor: Goldberg, Ken.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2025.
- 초록/해제
- 요약Robot learning has made significant strides in enabling complex manipulation skills across a variety of tasks. Yet, a central challenge remains: achieving generalization. Current methods often falter when faced with unseen objects, novel embodiments, or varying environments. This dissertation addresses this challenge through data-centric approaches to enhance the generalization of robot manipulation policies, focusing on three pillars: the combined use of simulation and real-world data, training-time data augmentation, and test-time adaptation. The core thesis is that strategically leveraging and augmenting data leads to robot systems that are more data-efficient, robust, and adaptable.The first two parts investigate training-time strategies using simulation and data augmentation to build robustness into the policy. This includes bridging the reality gap through self-supervision and co-training across simulated and real-world domains, augmenting data to simulate varied robot morphologies and camera viewpoints, and leveraging simple single-arm demonstrations to train coordinated bimanual policies.The third part introduces test-time adaptation techniques that avoid costly retraining. I present methods for adapting to new objects, transferring policies across diverse robot arms, rapidly imitating novel tasks with transformers, and enabling zero-shot grasping via multi-modal models-allowing robots to handle novelty in tasks, objects, and embodiments.Through extensive simulation and real-world experiments, this dissertation shows that data-centric methods-simulation, augmentation, and adaptation-can enhance generalization across tasks, objects, and embodiments, while making robot learning more robust and efficient.
- 일반주제명
- Robotics
- 일반주제명
- Computer science
- 일반주제명
- Industrial engineering
- 키워드
- Robot learning
- 키워드
- Robot systems
- 기타저자
- University of California, Berkeley Industrial Engineering & Operations Research
- 기본자료저록
- Dissertations Abstracts International. 87-01B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357861
■00520260202103610
■006m o d
■007cr#unu||||||||
■020 ▼a9798288865503
■035 ▼a(MiAaPQ)AAI32043083
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a629.8
■1001 ▼aChen, Lawrence Yunliang.
■24510▼aData-Centric Generalization for Robot Manipulation: Simulation, Augmentation, and Adaptation
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a243 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: B.
■500 ▼aAdvisor: Goldberg, Ken.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2025.
■520 ▼aRobot learning has made significant strides in enabling complex manipulation skills across a variety of tasks. Yet, a central challenge remains: achieving generalization. Current methods often falter when faced with unseen objects, novel embodiments, or varying environments. This dissertation addresses this challenge through data-centric approaches to enhance the generalization of robot manipulation policies, focusing on three pillars: the combined use of simulation and real-world data, training-time data augmentation, and test-time adaptation. The core thesis is that strategically leveraging and augmenting data leads to robot systems that are more data-efficient, robust, and adaptable.The first two parts investigate training-time strategies using simulation and data augmentation to build robustness into the policy. This includes bridging the reality gap through self-supervision and co-training across simulated and real-world domains, augmenting data to simulate varied robot morphologies and camera viewpoints, and leveraging simple single-arm demonstrations to train coordinated bimanual policies.The third part introduces test-time adaptation techniques that avoid costly retraining. I present methods for adapting to new objects, transferring policies across diverse robot arms, rapidly imitating novel tasks with transformers, and enabling zero-shot grasping via multi-modal models-allowing robots to handle novelty in tasks, objects, and embodiments.Through extensive simulation and real-world experiments, this dissertation shows that data-centric methods-simulation, augmentation, and adaptation-can enhance generalization across tasks, objects, and embodiments, while making robot learning more robust and efficient.
■590 ▼aSchool code: 0028.
■650 4▼aRobotics
■650 4▼aComputer science
■650 4▼aIndustrial engineering
■653 ▼aRobot learning
■653 ▼aRobot systems
■653 ▼aData augmentation
■653 ▼aData-centric methods
■690 ▼a0771
■690 ▼a0800
■690 ▼a0984
■690 ▼a0796
■690 ▼a0546
■71020▼aUniversity of California, Berkeley▼bIndustrial Engineering & Operations Research.
■7730 ▼tDissertations Abstracts International▼g87-01B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357861▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


