본문

서브메뉴

Embodied, Reflexive, and Multimodal Intelligence for Manipulation in Unstructured Environments
Embodied, Reflexive, and Multimodal Intelligence for Manipulation in Unstructured Environm...
Embodied, Reflexive, and Multimodal Intelligence for Manipulation in Unstructured Environments

상세정보

자료유형  
 학위논문 서양
최종처리일시  
20260202105613
ISBN  
9798265426949
DDC  
500
저자명  
Brouwer, Dane.
서명/저자  
Embodied, Reflexive, and Multimodal Intelligence for Manipulation in Unstructured Environments
발행사항  
[Sl] : Stanford University, 2025
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2025
형태사항  
157 p
주기사항  
Source: Dissertations Abstracts International, Volume: 87-06, Section: B.
주기사항  
Advisor: Cutkosky, Mark.
학위논문주기  
Thesis (Ph.D.)--Stanford University, 2025.
초록/해제  
요약Robotic manipulation in unstructured environments presents critical challenges that conventional industrial robots-typically rigid systems optimized for precision and repeatability-are ill-equipped to handle. Tasks such as retrieving objects from cluttered shelves or carefully handling delicate items amidst visual occlusion require careful mechanical design and more than just visual and proprioceptive sensor streams. To operate effectively in these environments, robots must embrace contact instead of avoiding it, and must intelligently integrate mechanical design, sensor feedback, and reactive behaviors. This work contributes three complementary strategies that leverage embodied intelligence, reflexive control, and multimodal machine learning to perform manipulation in complex real-world settings.First, we focus on how passive mechanical properties can embody robots with inherent capabilities for grasping highly variable objects. This type of intentional passive design in robotics is commonly called embodied intelligence. Toward this goal, we present farmHand , a semi-anthropomorphic robotic hand that extends gecko-inspired adhesive technology beyond its traditional manipulation limits by integrating it with novel compliant finger pads. These ribbed finger pads can maintain large contact area even under misalignment, provide an anti-peeling behavior, and enable load sharing across contact points through degressive shear stiffness. Empirical testing confirms farmHand 's key shape matching and load sharing characteristics in addition to demonstrating manipulation tasks across various object types and scales. This design utilizes gecko-inspired adhesion in tandem with embodied intelligence to bridge the performance gap between parallel jaw grippers and anthropomorphic systems, enabling robust grasping of objects with unpredictable geometries and material properties.Second, we develop a low-level, reactive control framework for navigating dense clutter. We propose two motion primitives-"burrow" and "excavate"-to mitigate jamming when reaching amidst constrained clutter. Experiments in hardware and simulation validate that the proposed reflexes, even when deployed with time-based triggers, enable robots to advance through dense configurations of movable objects, avoiding jamming and facilitating progress where straight line motions fail. We then develop a novel soft triaxial tactile sensor which is used to build a hybrid controller that utilizes event-based triggers to deploy the primitives. This event-based hybrid controller demonstrates over 80% success in hardware experiments, outperforming both straight line motions and time-based deployment of the primitives. These results underscore the importance of local force-based feedback for dynamic interaction, especially in severely occluded and cluttered environments.Finally, we investigate the impact of force sensing for high-level motion planning in unstructured environments. We conduct multimodal learning for contact-rich manipulation tasks, focusing on non-prehensile, visually occluded, and physically constrained object retraction. Using vision, proprioception, wrench information, soft tactile sensors, and suction grasping, our system learns to gently retract objects from dense clutter through imitation learning. An ablation experiment on a set of four trained policies-ranging from the baseline using vision and proprioception only to fully force-informed-reveal that tactile and wrench sensing greatly improve performance on this task. All policies with access to any force sensing modality improve safety, reduce completion time, and significantly improve success rate, with the fully force-informed policy outperforming the baseline policy by 80%. Each policy which lacked even a single force modality exhibited significantly increased timeout failures, indicating the importance of providing policies with observations that appropriately correspond with demonstrated strategies.Together, these contributions provide a roadmap for a new generation of robotic manipulators that combine mechanical intelligence, reactive control, and high-level learning informed by multimodal force information. The integration of thoughtful mechanical design, tactile sensing not just at the fingertips, and force-informed strategies at each level in the control architecture represents a critical shift in the design of robot systems capable of robustly interacting with complex physical environments.
일반주제명  
Kinematics
일반주제명  
Mechanical engineering
일반주제명  
Robotics
키워드  
Machine learning
키워드  
Robotic manipulation
기타저자  
Stanford University.
기본자료저록  
Dissertations Abstracts International. 87-06B.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008260126s2025        us                              c    eng  d
■001000017360740
■00520260202105613
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798265426949
■035    ▼a(MiAaPQ)AAI32316418
■035    ▼a(MiAaPQ)Stanfordrq555xv6342
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a500
■1001  ▼aBrouwer,  Dane.
■24510▼aEmbodied,  Reflexive,  and  Multimodal  Intelligence  for  Manipulation  in  Unstructured  Environments
■260    ▼a[Sl]▼bStanford  University▼c2025
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2025
■300    ▼a157  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  87-06,  Section:  B.
■500    ▼aAdvisor:  Cutkosky,  Mark.
■5021  ▼aThesis  (Ph.D.)--Stanford  University,  2025.
■520    ▼aRobotic  manipulation  in  unstructured  environments  presents  critical  challenges  that  conventional  industrial  robots-typically  rigid  systems  optimized  for  precision  and  repeatability-are  ill-equipped  to  handle.  Tasks  such  as  retrieving  objects  from  cluttered  shelves  or  carefully  handling  delicate  items  amidst  visual  occlusion  require  careful  mechanical  design  and  more  than  just  visual  and  proprioceptive  sensor  streams.  To  operate  effectively  in  these  environments,  robots  must  embrace  contact  instead  of  avoiding  it,  and  must  intelligently  integrate  mechanical  design,  sensor  feedback,  and  reactive  behaviors.  This  work  contributes  three  complementary  strategies  that  leverage  embodied  intelligence,  reflexive  control,  and  multimodal  machine  learning  to  perform  manipulation  in  complex  real-world  settings.First,  we  focus  on  how  passive  mechanical  properties  can  embody  robots  with  inherent  capabilities  for  grasping  highly  variable  objects.  This  type  of  intentional  passive  design  in  robotics  is  commonly  called  embodied  intelligence.  Toward  this  goal,  we  present  farmHand  ,  a  semi-anthropomorphic  robotic  hand  that  extends  gecko-inspired  adhesive  technology  beyond  its  traditional  manipulation  limits  by  integrating  it  with  novel  compliant  finger  pads.  These  ribbed  finger  pads  can  maintain  large  contact  area  even  under  misalignment,  provide  an  anti-peeling  behavior,  and  enable  load  sharing  across  contact  points  through  degressive  shear  stiffness.  Empirical  testing  confirms  farmHand  's  key  shape  matching  and  load  sharing  characteristics  in  addition  to  demonstrating  manipulation  tasks  across  various  object  types  and  scales.  This  design  utilizes  gecko-inspired  adhesion  in  tandem  with  embodied  intelligence  to  bridge  the  performance  gap  between  parallel  jaw  grippers  and  anthropomorphic  systems,  enabling  robust  grasping  of  objects  with  unpredictable  geometries  and  material  properties.Second,  we  develop  a  low-level,  reactive  control  framework  for  navigating  dense  clutter.  We  propose  two  motion  primitives-"burrow"  and  "excavate"-to  mitigate  jamming  when  reaching  amidst  constrained  clutter.  Experiments  in  hardware  and  simulation  validate  that  the  proposed  reflexes,  even  when  deployed  with  time-based  triggers,  enable  robots  to  advance  through  dense  configurations  of  movable  objects,  avoiding  jamming  and  facilitating  progress  where  straight  line  motions  fail.  We  then  develop  a  novel  soft  triaxial  tactile  sensor  which  is  used  to  build  a  hybrid  controller  that  utilizes  event-based  triggers  to  deploy  the  primitives.  This  event-based  hybrid  controller  demonstrates  over  80%  success  in  hardware  experiments,  outperforming  both  straight  line  motions  and  time-based  deployment  of  the  primitives.  These  results  underscore  the  importance  of  local  force-based  feedback  for  dynamic  interaction,  especially  in  severely  occluded  and  cluttered  environments.Finally,  we  investigate  the  impact  of  force  sensing  for  high-level  motion  planning  in  unstructured  environments.  We  conduct  multimodal  learning  for  contact-rich  manipulation  tasks,  focusing  on  non-prehensile,  visually  occluded,  and  physically  constrained  object  retraction.  Using  vision,  proprioception,  wrench  information,  soft  tactile  sensors,  and  suction  grasping,  our  system  learns  to  gently  retract  objects  from  dense  clutter  through  imitation  learning.  An  ablation  experiment  on  a  set  of  four  trained  policies-ranging  from  the  baseline  using  vision  and  proprioception  only  to  fully  force-informed-reveal  that  tactile  and  wrench  sensing  greatly  improve  performance  on  this  task.  All  policies  with  access  to  any  force  sensing  modality  improve  safety,  reduce  completion  time,  and  significantly  improve  success  rate,  with  the  fully  force-informed  policy  outperforming  the  baseline  policy  by  80%.  Each  policy  which  lacked  even  a  single  force  modality  exhibited  significantly  increased  timeout  failures,  indicating  the  importance  of  providing  policies  with  observations  that  appropriately  correspond  with  demonstrated  strategies.Together,  these  contributions  provide  a  roadmap  for  a  new  generation  of  robotic  manipulators  that  combine  mechanical  intelligence,  reactive  control,  and  high-level  learning  informed  by  multimodal  force  information.  The  integration  of  thoughtful  mechanical  design,  tactile  sensing  not  just  at  the  fingertips,  and  force-informed  strategies  at  each  level  in  the  control  architecture  represents  a  critical  shift  in  the  design  of  robot  systems  capable  of  robustly  interacting  with  complex  physical  environments.
■590    ▼aSchool  code:  0212.
■650  4▼aKinematics
■650  4▼aMechanical  engineering
■650  4▼aRobotics
■653    ▼aMachine  learning
■653    ▼aRobotic  manipulation
■690    ▼a0548
■690    ▼a0771
■71020▼aStanford  University.
■7730  ▼tDissertations  Abstracts  International▼g87-06B.
■790    ▼a0212
■791    ▼aPh.D.
■792    ▼a2025
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360740▼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
    TF17673 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.