서브메뉴
검색
Autonomous Methods for Learning and Pruning Motion Primitives for Navigation and Adversarial Tasks
Autonomous Methods for Learning and Pruning Motion Primitives for Navigation and Adversarial Tasks
Detailed Information
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105517
- ISBN
- 9798263345662
- DDC
- 330
- 서명/저자
- Autonomous Methods for Learning and Pruning Motion Primitives for Navigation and Adversarial Tasks
- 발행사항
- [Sl] : Georgia Institute of Technology, 2023
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2023
- 형태사항
- 112 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-05, Section: B.
- 주기사항
- Advisor: Mazumdar, Anirban.
- 학위논문주기
- Thesis (Ph.D.)--Georgia Institute of Technology, 2023.
- 초록/해제
- 요약Kinodynamic motion planning presents the difficult problem of finding a trajectory to reach a goal under both kinematic and dynamic constraints. Motion primitives can simplify this problem by solving the dynamics offline and allowing for fast planning via concatenation. Primitive-based planners offer great advantages in real-time applications where run-time is a critical concern. A key challenge with these methods is the design of a primitive library which concisely captures the capabilities of the system.This work explores the use of reinforcement learning and genetic algorithms to autonomously generate motion primitive libraries and prune them down to their most effective components. Additionally, we develop improvements for primitive-based motion planners for navigation and adversarial games. A number of simulations are performed using a six degree-of-freedom F-16 model to demonstrate the benefit of planning improvements and the efficacy of the learned motion primitives.This work begins by examining the problem of navigation with and without obstacles. We propose the use of Hybrid A* and discuss it's strengths and weaknesses compared to alternative algorithms. Additionally, we propose two enhancements to the algorithm to improve its performance. First, we present a set of heuristics acquired by pre-computed subproblems to reduce the runtime of the search. Second, we modify the algorithm with a post-process optimization step to improve the resolution of the path with minimal increase in runtime.The first core contribution of this work is a motion primitive learning for which we design special shaping rewards and an extraction algorithm. The shaping rewards encourage a Soft Actor-Critic agent to visit trim states while solving a given task. The extraction algorithm uses these trim states to identify valid maneuvers and add them to the primitive library. We demonstrate this framework on a navigation task by applying the learned motion primitives with Hybrid A. The results show substantial improvement over a manually designed base library, indicating that the framework is capable of generating effective maneuvers.Our second contribution is the formulation of a genetic algorithm to select the best subset of motion primitives from a given library. This algorithm balances the planning time and final cost to produce the smallest libraries with the best path quality. We propose a set of mutation operations to incrementally modify primitive libraries while maintaining reachability. We apply this algorithm to a learned primitive library and demonstrate its ability to create a several primitive libraries with different trade-offs for planning speed and performance.Finally, we extend this learning framework to an adversarial context which poses a more difficult and dynamic environment to generate effective motion primitives. We de-sign a primitive-based Monte Carlo Tree Search to apply motion primitives to adversarial tasks. Additionally, we explore the use of beam search to reduce planning time for practical applications. We apply the learning framework to an adversarial example and show that it produces effective motion primitives for the task. Additionally, we compare this approach to a forward simulated Monte Carlo Tree Search used in prior literature and discuss the strengths and weaknesses of each.
- 일반주제명
- Aircraft
- 일반주제명
- Kinematics
- 일반주제명
- Mutation
- 일반주제명
- Planning
- 일반주제명
- Genetic algorithms
- 일반주제명
- Real time
- 일반주제명
- Symmetry
- 일반주제명
- Design
- 일반주제명
- Libraries
- 기본자료저록
- Dissertations Abstracts International. 87-05B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2023 us c eng d■001000017360391
■00520260202105517
■006m o d
■007cr#unu||||||||
■020 ▼a9798263345662
■035 ▼a(MiAaPQ)AAI32309377
■035 ▼a(MiAaPQ)GeorgiaTech75540
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a330
■1001 ▼aGoddard, Zachary.
■24510▼aAutonomous Methods for Learning and Pruning Motion Primitives for Navigation and Adversarial Tasks
■260 ▼a[Sl]▼bGeorgia Institute of Technology▼c2023
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2023
■300 ▼a112 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-05, Section: B.
■500 ▼aAdvisor: Mazumdar, Anirban.
■5021 ▼aThesis (Ph.D.)--Georgia Institute of Technology, 2023.
■520 ▼aKinodynamic motion planning presents the difficult problem of finding a trajectory to reach a goal under both kinematic and dynamic constraints. Motion primitives can simplify this problem by solving the dynamics offline and allowing for fast planning via concatenation. Primitive-based planners offer great advantages in real-time applications where run-time is a critical concern. A key challenge with these methods is the design of a primitive library which concisely captures the capabilities of the system.This work explores the use of reinforcement learning and genetic algorithms to autonomously generate motion primitive libraries and prune them down to their most effective components. Additionally, we develop improvements for primitive-based motion planners for navigation and adversarial games. A number of simulations are performed using a six degree-of-freedom F-16 model to demonstrate the benefit of planning improvements and the efficacy of the learned motion primitives.This work begins by examining the problem of navigation with and without obstacles. We propose the use of Hybrid A* and discuss it's strengths and weaknesses compared to alternative algorithms. Additionally, we propose two enhancements to the algorithm to improve its performance. First, we present a set of heuristics acquired by pre-computed subproblems to reduce the runtime of the search. Second, we modify the algorithm with a post-process optimization step to improve the resolution of the path with minimal increase in runtime.The first core contribution of this work is a motion primitive learning for which we design special shaping rewards and an extraction algorithm. The shaping rewards encourage a Soft Actor-Critic agent to visit trim states while solving a given task. The extraction algorithm uses these trim states to identify valid maneuvers and add them to the primitive library. We demonstrate this framework on a navigation task by applying the learned motion primitives with Hybrid A. The results show substantial improvement over a manually designed base library, indicating that the framework is capable of generating effective maneuvers.Our second contribution is the formulation of a genetic algorithm to select the best subset of motion primitives from a given library. This algorithm balances the planning time and final cost to produce the smallest libraries with the best path quality. We propose a set of mutation operations to incrementally modify primitive libraries while maintaining reachability. We apply this algorithm to a learned primitive library and demonstrate its ability to create a several primitive libraries with different trade-offs for planning speed and performance.Finally, we extend this learning framework to an adversarial context which poses a more difficult and dynamic environment to generate effective motion primitives. We de-sign a primitive-based Monte Carlo Tree Search to apply motion primitives to adversarial tasks. Additionally, we explore the use of beam search to reduce planning time for practical applications. We apply the learning framework to an adversarial example and show that it produces effective motion primitives for the task. Additionally, we compare this approach to a forward simulated Monte Carlo Tree Search used in prior literature and discuss the strengths and weaknesses of each.
■590 ▼aSchool code: 0078.
■650 4▼aAircraft
■650 4▼aKinematics
■650 4▼aMutation
■650 4▼aPlanning
■650 4▼aGenetic algorithms
■650 4▼aReal time
■650 4▼aSymmetry
■650 4▼aDesign
■650 4▼aLibraries
■690 ▼a0389
■690 ▼a0800
■71020▼aGeorgia Institute of Technology.
■7730 ▼tDissertations Abstracts International▼g87-05B.
■790 ▼a0078
■791 ▼aPh.D.
■792 ▼a2023
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360391▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
Preview
Export
ChatGPT Discussion
AI Recommended Related Books
Подробнее информация.
- Бронирование
- не существует
- моя папка
- Первый запрос зрения
- Non-Book Loan Application
- Nighttime Book Loan Application
Available after logging in.


