서브메뉴
검색
Optimization-Driven Autonomous Driving: Control, Intent, and Testing
Optimization-Driven Autonomous Driving: Control, Intent, and Testing
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152748
- ISBN
- 9798342107457
- DDC
- 320.52
- 저자명
- Dyro, Robert.
- 서명/저자
- Optimization-Driven Autonomous Driving: Control, Intent, and Testing
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 129 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-04, Section: A.
- 주기사항
- Advisor: Pavone, Marco.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약The primary goal of this research is to develop behavior models amenable to the task of online intent inference. This is crucial as real-time intent inference is increasingly essential in dynamic environments where autonomous systems interact with humans as the number of other traffic participants increases. Secondly, building on developments in behavior models, we also focus on better Autonomous Vehicles (AV) testingvia counterfactual scenario creation, which is vital for enhancing the safety and reliability of AVs in complex and unpredictable real-world scenarios. We develop new optimization tools for accelerating the solution of bilevel programs, targeting online intent inference as a primary application.The methodological contribution leads us to focus on one type of behavior model: an inverse optimal control-based (IOC) agent behavior model. Our methodology removes the typical obstacle of applying IOC online, which is a large number of costly iterations, thus enabling online intent inference. We combine our new methodology with previous IOC approaches, leading to a hybrid method. This innovative approach addresses the urgent need for adaptable and efficient online intent inference in rapidly changing scenarios. We do so to allow both fast online intent inference in structured models with comparatively few but interpretable parameters (e.g., parameterized rule-based models, collision repulsion parameterization) and unstructured, expressive models with up to billions of parameters (e.g., neural network residual corrector models). This allows our methodology to speed up existing largely structured IOC approaches while also taking advantage of the power of unstructured deep learning.Acknowledging the growing public concern about the safety of AVs, we attempt to develop better testing for AVs. To do so, we leverage our improved behavior modeling techniques in conjunction with recent advancements in behavior models. We aim to facilitate more exhaustive AV testing by generating realistic, counterfactual scenarios, allowing for stress testing of AVs. Our method relies on extracting behavior probability distribution encoded in fitted behavior models to generate challenging but realistic counterfactual scenarios automatically. Our method has the advantage of being generally applicable to parametric behavior models, generating a diversity of scenarios and allowing for the explicit specification of the lower probability limit of the generated scenarios - blending stress-testing with realism.
- 일반주제명
- Conservatism
- 일반주제명
- Robust control
- 일반주제명
- Deep learning
- 일반주제명
- Sensitivity analysis
- 일반주제명
- Computer vision
- 일반주제명
- Planning
- 일반주제명
- Neural networks
- 일반주제명
- Decision making
- 일반주제명
- Convex analysis
- 일반주제명
- Realism
- 일반주제명
- Robotics
- 일반주제명
- Industrial engineering
- 일반주제명
- Mathematics
- 일반주제명
- Political science
- 일반주제명
- Computer science
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-04A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017163746
■00520250211152748
■006m o d
■007cr#unu||||||||
■020 ▼a9798342107457
■035 ▼a(MiAaPQ)AAI31520321
■035 ▼a(MiAaPQ)Stanfordts345km5140
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a320.52
■1001 ▼aDyro, Robert.
■24510▼aOptimization-Driven Autonomous Driving: Control, Intent, and Testing
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a129 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-04, Section: A.
■500 ▼aAdvisor: Pavone, Marco.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aThe primary goal of this research is to develop behavior models amenable to the task of online intent inference. This is crucial as real-time intent inference is increasingly essential in dynamic environments where autonomous systems interact with humans as the number of other traffic participants increases. Secondly, building on developments in behavior models, we also focus on better Autonomous Vehicles (AV) testingvia counterfactual scenario creation, which is vital for enhancing the safety and reliability of AVs in complex and unpredictable real-world scenarios. We develop new optimization tools for accelerating the solution of bilevel programs, targeting online intent inference as a primary application.The methodological contribution leads us to focus on one type of behavior model: an inverse optimal control-based (IOC) agent behavior model. Our methodology removes the typical obstacle of applying IOC online, which is a large number of costly iterations, thus enabling online intent inference. We combine our new methodology with previous IOC approaches, leading to a hybrid method. This innovative approach addresses the urgent need for adaptable and efficient online intent inference in rapidly changing scenarios. We do so to allow both fast online intent inference in structured models with comparatively few but interpretable parameters (e.g., parameterized rule-based models, collision repulsion parameterization) and unstructured, expressive models with up to billions of parameters (e.g., neural network residual corrector models). This allows our methodology to speed up existing largely structured IOC approaches while also taking advantage of the power of unstructured deep learning.Acknowledging the growing public concern about the safety of AVs, we attempt to develop better testing for AVs. To do so, we leverage our improved behavior modeling techniques in conjunction with recent advancements in behavior models. We aim to facilitate more exhaustive AV testing by generating realistic, counterfactual scenarios, allowing for stress testing of AVs. Our method relies on extracting behavior probability distribution encoded in fitted behavior models to generate challenging but realistic counterfactual scenarios automatically. Our method has the advantage of being generally applicable to parametric behavior models, generating a diversity of scenarios and allowing for the explicit specification of the lower probability limit of the generated scenarios - blending stress-testing with realism.
■590 ▼aSchool code: 0212.
■650 4▼aConservatism
■650 4▼aRobust control
■650 4▼aDeep learning
■650 4▼aSensitivity analysis
■650 4▼aComputer vision
■650 4▼aPlanning
■650 4▼aNeural networks
■650 4▼aDecision making
■650 4▼aConvex analysis
■650 4▼aRealism
■650 4▼aRobotics
■650 4▼aIndustrial engineering
■650 4▼aMathematics
■650 4▼aPolitical science
■650 4▼aComputer science
■690 ▼a0771
■690 ▼a0800
■690 ▼a0546
■690 ▼a0405
■690 ▼a0615
■690 ▼a0984
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-04A.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17163746▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


