Detail View

A Study on Reinforcement Learning-Driven Snake Robot

Citations

WEB OF SCIENCE

Citations

SCOPUS

Metadata Downloads

DC Field Value Language
dc.contributor.advisor 윤동원 -
dc.contributor.author Bongsub Song -
dc.date.accessioned 2026-01-23T10:59:10Z -
dc.date.available 2026-01-23T10:59:10Z -
dc.date.issued 2025 -
dc.identifier.uri https://scholar.dgist.ac.kr/handle/20.500.11750/59771 -
dc.identifier.uri http://dgist.dcollection.net/common/orgView/200000891302 -
dc.description Snake robot, Deep reinforcement learning, Bio-inspired gait decomposition, Real-world application -
dc.description.abstract Snake robots, inspired by the complex and adaptable locomotion of natural snakes, present significant potential in various applications ranging from search-and-rescue operations to medical procedures. However, their high degree of freedom makes their control a challenging task, often hindering their wider adoption. This study introduces a novel approach to addressing this issue: the application of deep reinforcement learning (DRL) to handle the intricate dynamics of snake robots.
This research presents a comprehensive review of snake robots, their anatomy, gaits, and previous models. It delves into function approximation techniques, with a particular focus on neural networks. Further, it explores the fundamentals of reinforcement learning, including policy and value-based methods and established algorithms.
Leveraging this theoretical foundation, we design and develop a modular snake robot, implement a control framework, and create a reinforcement learning environment. We apply DRL to decompose the gaits of the snake robot, estimate its dynamics, and facilitate smooth gait changes through multi-agent learning.
The developed snake robot will be tested in real-world applications, showcasing the effectiveness of our approach. This research not only propels the field of snake robot research but also brings us a step closer to the broader adoption of these flexible and versatile robots in real-world applications.|뱀 로봇은 자연계 뱀의 복잡하고 유연한 움직임에서 영감을 받아 개발된 로봇으로, 수색 및 구조 작업부터 의료 시술에 이르기까지 다양한 응용 가능성을 지니고 있다. 그러나 뱀 로봇의 높은 자유도는 제어를 어렵게 만들며, 이는 그 활용 범위를 제한하는 주요 요인 중 하나이다. 본 연구는 이러한 제어 문제를 해결하기 위한 새로운 접근으로, 심층 강화학습(Deep Reinforcement Learning, DRL)을 적용하여 뱀 로봇의 복잡한 동역학을 효과적으로 다루는 방법을 제시한다.

연구 초기에는 뱀 로봇의 해부학적 구조, 다양한 Gait, 기존 모델에 대한 종합적인 문헌 조사를 수행하였다. 이후 함수 근사 기법, 특히 신경망 기반의 접근 방식과 함께 강화학습의 기본 개념(정책 기반 및 가치 기반 학습, 주요 알고리즘 등)을 이론적으로 고찰하였다.

이러한 이론적 토대를 바탕으로, 본 연구에서는 모듈형 뱀 로봇을 설계 및 제작하고, 제어 프레임워크를 구현하였으며, 이를 위한 강화학습 환경을 구축하였다. 제안된 방법은 뱀 로봇의 Gait를 분해하고, 그 동역학을 추정하며, 다중 에이전트 학습을 통해 Gait 전이를 매끄럽게 수행할 수 있도록 한다.

개발된 뱀 로봇은 향후 실제 환경에서의 실험을 통해 그 성능이 검증될 예정이며, 본 연구는 뱀 로봇 제어 분야의 기술적 발전을 촉진함과 동시에, 유연하고 적응력 있는 로봇 시스템의 실용화를 앞당기는 데 기여할 것으로 기대된다.
-
dc.description.tableofcontents Ⅰ. Introduction 1
1.1 Motivation 1
1.2 Thesis approach 3
1.3 Thesis outline 5

II. Literature Study & Backgrounds 7
2.1 On snake robots 7
2.1.1 Biological snakes 9
2.1.2 Snake’s gaits 12
2.1.3 Prior Developments in Snake Robots 15
2.1.4 Modeling and analysis of snake robots 23
2.2 On function approximation 34
2.2.1 Linear regression 36
2.2.2 Logistic and multinomial regression 38
2.2.3 Neural networks 39
2.3 On reinforcement learning 44
2.3.1 Remarks and notations 47

IIⅠ. Development of Snake Robot 52
3.1 Overview of Developed Snake Robots 52
3.2 Snake robot A 55

3.2.1 Hardware design 55
3.2.2 Software design 57
3.3 Snake robot B 58
3.3.1 Hardware design 58
3.3.2 Sensor Integration 59
3.3.3 Electrical and Communication Structure 60

ⅠV. Optimizing Snake Robot Locomotion with Decomposed Gait Pattern Representation 64
4.1 Introduction 65
4.2 Gait decomposition 67
4.2.1 Serpenoid curve function 68
4.2.2 Motion matrix 70
4.2.3 Motor control algorithm 77
4.3 Gait parameter gradient 82
4.3.1 Dynamics model of snake robot 83
4.3.2 Robot configuration and simulation environment 85
4.3.3 Object function definition and optimization criteria 90
4.3.4 Gait parameter optimization process 96
4.4 Optimization results and analysis 114
4.4.1 Feasibility analysis of motion matrix optimization 114
4.4.2 Sensitivity analysis of utility function weights 117
4.4.3 Optimization results 122
4.5 Experiment 125
4.6 Discussion 131
4.7 Conclusion 131

V. Biologically Inspired Decomposed Gait Pattern Representation for Exploration Space Reduction in Reinforcement Learning of Snake Robot 134
5.1 Introduction 135
5.2 Design of the snake robot 139
5.2.1 Link module structure for ground contact sensing 143
5.2.2 System architecture and operational structure 144
5.3 Torque-based gait decomposition 148
5.3.1 Motion matrix transformation 149
5.3.2 Gait generation algorithm 162
5.4 Gait policy learning with reinforcement learning based approach 163
5.4.1 State, Action and Reward space in RL framework 164
5.4.2 Reward function design for gait optimization 165
5.4.3 Learning algorithms and training setup 167
5.4.4 Learning hyperparameters 168
5.4.5 Training results of gait policy 171
5.4.6 Discussion on learning results 177
5.5 Experiment setup and results 180
5.5.1 Experimental results and analysis 182
5.6 Conclusion 184

VⅠ. Future work 186

VII. Conclusion 188

Reference 191
-
dc.format.extent 204 -
dc.language eng -
dc.publisher DGIST -
dc.title A Study on Reinforcement Learning-Driven Snake Robot -
dc.title.alternative 심층 강화학습 기반의 뱀 로봇에 관한 연구 -
dc.type Thesis -
dc.identifier.doi 10.22677/THESIS.200000891302 -
dc.description.degree Doctor -
dc.contributor.department Department of Robotics and Mechatronics Engineering -
dc.contributor.coadvisor Jongwon Park -
dc.date.awarded 2025-08-01 -
dc.publisher.location Daegu -
dc.description.database dCollection -
dc.citation XT.RD 송45 202508 -
dc.date.accepted 2025-07-21 -
dc.contributor.alternativeDepartment 로봇및기계전자공학과 -
dc.subject.keyword Snake robot, Deep reinforcement learning, Bio-inspired gait decomposition, Real-world application -
dc.contributor.affiliatedAuthor Bongsub Song -
dc.contributor.affiliatedAuthor Dongwon Yun -
dc.contributor.affiliatedAuthor Jongwon Park -
dc.contributor.alternativeName 송봉섭 -
dc.contributor.alternativeName Dongwon Yun -
dc.contributor.alternativeName 박종원 -
dc.rights.embargoReleaseDate 2028-08-31 -
Show Simple Item Record

File Downloads

  • There are no files associated with this item.

공유

qrcode
공유하기

Total Views & Downloads

???jsp.display-item.statistics.view???: , ???jsp.display-item.statistics.download???: