서브메뉴
검색
Data-Driven Methods for Real-Time Control in Autonomous Racing Games
Data-Driven Methods for Real-Time Control in Autonomous Racing Games
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211151346
- ISBN
- 9798384448204
- DDC
- 621
- 서명/저자
- Data-Driven Methods for Real-Time Control in Autonomous Racing Games
- 발행사항
- [Sl] : University of California, Berkeley, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 124 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
- 주기사항
- Advisor: Tomizuka, Masayoshi.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2024.
- 초록/해제
- 요약Offline datasets and demonstrations can provide valuable guidance for planning and control in robotic and intelligent systems. Often, demonstration data consists of sequences of observations and control actions that occurred in the environment, demonstrating a specific task or desirable behavior. They can originate from human experts or demonstrators, simple rule-based or model-based controllers, or even previously machine-learned policies. However, this data is most commonly unstructured, without annotations of underlying tasks, rewards, thoughts, rules, or states that were used by the demonstration generator. For long-horizon, complicated tasks, these unstructured demonstrations often fail to provide enough information to facilitate learning. Thus, additional structure in sequential demonstrations could inform or enhance learning-based planning and control.This dissertation explores how meaningful partitions of sequential data can guide and accelerate learning, and improve the performance and reliability of control systems. In particular, we primarily focus on the complicated task of planning and control for an autonomous racing vehicle, which must rapidly traverse a closed racetrack with complicated, unknown vehicle dynamics. We identify three challenges that exist in current algorithms for autonomous racing systems. First, the control system should be robust to state deviations and maintain control when it encounters previously unseen states. Second, training the control system should efficiently leverage its knowledge to minimize the number of environment interactions necessary to train a high-performance system. Third, the control system should emulate human behavior to race in predictable and understandable ways. In this dissertation, we explore how offline datasets can address these challenges while maintaining good racing performance through both modular planning and control and end-to-end learning. In the first part, we explore control systems that are built on separate planning and control modules. In particular, we design a trajectory generation module that can leverage an existing demonstration for fast, online trajectory planning. In Chapter 2, we propose a novel acceleration motion primitive for online trajectory generation that can plan trajectories that are better suited for the complex racing environment than existing methods. Analysis of the trajectories reveals that our method generates trajectories reduce the error between generated trajectories and the offline demonstration and lead to less aggressive acceleration and jerk than existing velocity motion primitives. In Chapter 3, we learn a sequence of motion primitives from a reference trajectory, and use the primitive sequence to generate racing trajectories in real-time. When used with Model Predictive Control (MPC), our proposed acceleration motion primitive and learning-based trajectory generation algorithm allows the MPC to recover from deviations from the reference while maintaining racing speeds. In the second part, we explore how offline data can improve end-to-end control policy learning. In Chapter 4, we tackle the problem of learning a control policy from long horizon sparse rewards by adding additional hierarchical structure from offline data. Our Skill-Critic method for online fine-tuning of the hierarchical policies reduces the requirements for online interactions with the environment and results in the fastest racer. In Chapter 5, rather than learning from a designed reward function, we tackle the problem of imitating human driver demonstrations. We propose a method that leverages offline sequence modeling architectures and online fine-tuning to imitate human racers. Our proposed method is more sample efficient. The resulting policies are more stable and are able to achieve the fastest lap-times compared to existing methods.
- 일반주제명
- Mechanical engineering
- 일반주제명
- Computer engineering
- 일반주제명
- Automotive engineering
- 일반주제명
- Information technology
- 일반주제명
- Robotics
- 키워드
- Control systems
- 기타저자
- University of California, Berkeley Mechanical Engineering
- 기본자료저록
- Dissertations Abstracts International. 86-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017161363
■00520250211151346
■006m o d
■007cr#unu||||||||
■020 ▼a9798384448204
■035 ▼a(MiAaPQ)AAI31242663
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621
■1001 ▼aWeaver, Catherine Nichole.
■24510▼aData-Driven Methods for Real-Time Control in Autonomous Racing Games
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a124 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-03, Section: B.
■500 ▼aAdvisor: Tomizuka, Masayoshi.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2024.
■520 ▼aOffline datasets and demonstrations can provide valuable guidance for planning and control in robotic and intelligent systems. Often, demonstration data consists of sequences of observations and control actions that occurred in the environment, demonstrating a specific task or desirable behavior. They can originate from human experts or demonstrators, simple rule-based or model-based controllers, or even previously machine-learned policies. However, this data is most commonly unstructured, without annotations of underlying tasks, rewards, thoughts, rules, or states that were used by the demonstration generator. For long-horizon, complicated tasks, these unstructured demonstrations often fail to provide enough information to facilitate learning. Thus, additional structure in sequential demonstrations could inform or enhance learning-based planning and control.This dissertation explores how meaningful partitions of sequential data can guide and accelerate learning, and improve the performance and reliability of control systems. In particular, we primarily focus on the complicated task of planning and control for an autonomous racing vehicle, which must rapidly traverse a closed racetrack with complicated, unknown vehicle dynamics. We identify three challenges that exist in current algorithms for autonomous racing systems. First, the control system should be robust to state deviations and maintain control when it encounters previously unseen states. Second, training the control system should efficiently leverage its knowledge to minimize the number of environment interactions necessary to train a high-performance system. Third, the control system should emulate human behavior to race in predictable and understandable ways. In this dissertation, we explore how offline datasets can address these challenges while maintaining good racing performance through both modular planning and control and end-to-end learning. In the first part, we explore control systems that are built on separate planning and control modules. In particular, we design a trajectory generation module that can leverage an existing demonstration for fast, online trajectory planning. In Chapter 2, we propose a novel acceleration motion primitive for online trajectory generation that can plan trajectories that are better suited for the complex racing environment than existing methods. Analysis of the trajectories reveals that our method generates trajectories reduce the error between generated trajectories and the offline demonstration and lead to less aggressive acceleration and jerk than existing velocity motion primitives. In Chapter 3, we learn a sequence of motion primitives from a reference trajectory, and use the primitive sequence to generate racing trajectories in real-time. When used with Model Predictive Control (MPC), our proposed acceleration motion primitive and learning-based trajectory generation algorithm allows the MPC to recover from deviations from the reference while maintaining racing speeds. In the second part, we explore how offline data can improve end-to-end control policy learning. In Chapter 4, we tackle the problem of learning a control policy from long horizon sparse rewards by adding additional hierarchical structure from offline data. Our Skill-Critic method for online fine-tuning of the hierarchical policies reduces the requirements for online interactions with the environment and results in the fastest racer. In Chapter 5, rather than learning from a designed reward function, we tackle the problem of imitating human driver demonstrations. We propose a method that leverages offline sequence modeling architectures and online fine-tuning to imitate human racers. Our proposed method is more sample efficient. The resulting policies are more stable and are able to achieve the fastest lap-times compared to existing methods.
■590 ▼aSchool code: 0028.
■650 4▼aMechanical engineering
■650 4▼aComputer engineering
■650 4▼aAutomotive engineering
■650 4▼aInformation technology
■650 4▼aRobotics
■653 ▼aAutonomous racing
■653 ▼aAutonomous vehicles
■653 ▼aControl systems
■653 ▼aImitation learning
■653 ▼aReinforcement learning
■690 ▼a0548
■690 ▼a0489
■690 ▼a0464
■690 ▼a0800
■690 ▼a0771
■690 ▼a0540
■71020▼aUniversity of California, Berkeley▼bMechanical Engineering.
■7730 ▼tDissertations Abstracts International▼g86-03B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17161363▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


