Detail View

Control-Informed Networking: Designing Cyber-Physical Systems Networks with Integrated Control

Citations

WEB OF SCIENCE

Citations

SCOPUS

Metadata Downloads

Title
Control-Informed Networking: Designing Cyber-Physical Systems Networks with Integrated Control
Alternative Title
제어 인지 네트워킹: 통합 제어를 고려한 사이버-물리 시스템 네트워크 설계
DGIST Authors
Hyung-Seok ParkKyung-Joon ParkYongsoon Eun
Advisor
박경준
Co-Advisor(s)
Yongsoon Eun
Issued Date
2026
Awarded Date
2026-08-01
Type
Thesis
Description
Cyber-physical systems, parameter controllability, ROS 2 DDS, latency modeling, industrial wireless networks, multi-armed bandit, control-informed networking
Abstract

As robotic systems are increasingly deployed in real-world industrial environments, their dependence on wireless networks for communication with fleet management systems, remote operators, and cooperating robots has become a critical performance bottleneck. Yet the robotic systems community has largely treated the network as a transparent pipe, lacking quantitative understanding of how network parameters affect communication behavior and control-relevant performance. This dissertation addresses this gap by analyzing network parameters in robotic cyber-physical systems (CPS), identifying which parameters can be configured by the designer and which are imposed by the wireless environment, and optimizing each category through the appropriate mechanism.
We begin by deriving a closed-form latency model L(p,h,r) for the Data Distribution Service (DDS) in ROS 2, expressing the average communication latency as a function of the packet delivery ratio p, the heartbeat period h, and the data period r. Validated with an average error of 6.88% across 35 experimental scenarios, the model reveals that all three parameters significantly influence latency, but their natures differ fundamentally: h and r are design variables that the engineer can directly configure and optimize, whereas p is an environment variable determined by the wireless channel that cannot be set by design. Moreover, changes in p have a large and nonlinear effect: a 21% reduction in p causes up to 13.2x increase in latency, meaning that even after configuring h and r, the system remains vulnerable to the wireless environment. This finding establishes that design variable optimization is necessary but insufficient, motivating a complementary mechanism to discover the best available p.
To address this remaining vulnerability, we propose the Criticality-Aware Adaptive Path Learning (CAPL) algorithm for industrial wireless sensor-actuator networks (IWSAN). In a multi-path network, different paths offer different values of p, but the best path is initially unknown and must be discovered through exploration. This exploration poses a fundamental dilemma in CPS: transmitting on an untested path risks losing a control packet, and the physical consequences of that loss depend on the current state of the controlled system. CAPL resolves this dilemma by introducing a physically-grounded concept of dynamic packet criticality, using low-criticality moments, when the system can tolerate packet loss, for network exploration. Simulation and testbed experiments demonstrate that CAPL reduces the average settling time by 18.25%, 7.19%, and 35.58% compared to epsilon-greedy, SWUCB, and EXP3, respectively.
These two contributions, analysis at the application layer (DDS/Ethernet) and search at the routing layer (IEEE 802.15.4/IWSAN), are abstracted into a control-informed networking framework based on parameter controllability. The framework follows an analyze-classify-act structure: model how network parameters affect performance, identify which parameters are configurable or environmental, optimize the design variables analytically, and discover the best available environment variable through control-state-guided exploration. The cross-layer, cross-stack evidence suggests that this perspective is useful beyond either individual protocol in a limited but practical sense. Together, the framework shows that effective CPS network design requires both analytical understanding of configurable parameters and control-state-guided search for environmental parameters.|로봇 시스템이 물류, 제조, 서비스 등 실제 산업 환경에 광범위하게 배치됨에 따라, 플릿 관리 시스템, 원격 운영자, 협력 로봇과의 무선 네트워크 통신 의존도가 핵심 성능 병목으로 부상하고 있다. 그러나 로봇 시스템 커뮤니티는 네트워크를 투명한 파이프로 간주해 왔으며, 네트워크 파라미터가 제어 성능에 미치는 영향에 대한 정량적 이해가 부족한 상황이다. 본 학위논문은 로봇 사이버-물리 시스템(CPS)의 네트워크 파라미터를 분석하고, 설계자가 직접 설정할 수 있는 파라미터와 무선 환경에 의해 결정되는 파라미터를 구분한 뒤, 각 범주에 적합한 방식으로 최적화하는 제어 인지 네트워킹 관점을 제시한다.
먼저 ROS 2의 데이터 분산 서비스(DDS)에 대한 폐쇄형 지연 모델 L(p,h,r)을 도출하여, 평균 통신 지연을 패킷 전달률 p, 하트비트 주기 h, 데이터 주기 r의 함수로 표현한다. 35개 실험 시나리오에 걸쳐 평균 오차 6.88%로 검증된 이 모델은 세 파라미터 모두 지연에 유의미한 영향을 미치지만 그 성격이 근본적으로 다름을 밝힌다: h와 r은 엔지니어가 직접 설정하고 최적화할 수 있는 설계 변수(design variable)인 반면, p는 무선 채널에 의해 결정되어 설계로 제어할 수 없는 환경 변수(environment variable)이다. 특히 p의 변화는 크고 비선형적인 영향을 미쳐, p가 21% 감소하면 지연이 최대 13.2배 증가한다. 이는 설계 변수 h와 r을 조정한 이후에도 시스템 성능이 무선 환경에 취약할 수 있으며, 가능한 환경 중 더 나은 p를 발견하기 위한 보완적 메커니즘이 필요함을 의미한다.
이 취약성을 해소하기 위해 산업용 무선 센서-액추에이터 네트워크(IWSAN)를 위한 임계도 인지 적응형 경로 학습(CAPL) 알고리즘을 제안한다. 다중 경로 네트워크에서 서로 다른 경로는 서로 다른 p 값을 제공하지만, 최적 경로는 사전에 알 수 없으며 탐색을 통해 발견해야 한다. 이 탐색은 CPS에서 근본적인 딜레마를 야기한다: 미탐색 경로로의 전송은 제어 패킷 손실 위험을 수반하며, 그 물리적 결과는 피제어 시스템의 현재 상태에 의존한다. CAPL은 물리적 기반의 동적 패킷 임계도 개념을 도입하여, 시스템이 패킷 손실을 허용할 수 있는 저임계도 순간을 네트워크 탐색에 활용함으로써 이 딜레마를 해결한다. 시뮬레이션 및 테스트베드 실험에서 CAPL은 epsilon-greedy, SWUCB, EXP3 대비 평균 정착 시간을 각각 18.25%, 7.19%, 35.58% 감소시켰다.
응용 계층(DDS/Ethernet)에서의 분석과 라우팅 계층(IEEE 802.15.4/IWSAN)에서의 탐색이라는 두 기여를 파라미터 제어 가능성에 기반한 제어 인지 네트워킹 프레임워크로 추상화한다. 이 프레임워크는 CPS 네트워크 최적화를 분석-분류-실행 과정으로 구조화한다: 네트워크 파라미터가 성능에 미치는 영향을 모델링하고, 각 파라미터가 설정 가능한 설계 변수인지 환경 변수인지 구분한 뒤, 설계 변수는 해석적으로 최적화하고 환경 변수는 제어 상태 기반 탐색을 통해 가능한 최선의 값을 발견한다. 교차 계층, 교차 스택 증거는 이 관점이 특정 프로토콜 구현에만 묶이지 않는 유용한 설계 관점임을 보여준다.

더보기
Table Of Contents
1 Introduction 1
1.1 Robotic Systems and Network Dependence 1
1.2 The Networking Challenge in Robotic CPS 2
1.3 Research Questions and Approach 4
1.4 Contributions 5
1.5 Organization 7
2 Background and Related Work 8
2.1 Networked Cyber-Physical Systems 8
2.1.1 Fundamentals of CPS 8
2.1.2 Network-Mediated Communication in Robotic CPS 9
2.1.3 Emerging Wireless Technologies for Robotic CPS 10
2.1.4 Related Work: Networked Control Systems 11
2.2 ROS 2 and Data Distribution Service 12
2.2.1 Evolution from ROS 1 to ROS 2 12
2.2.2 Data-Centric Publish-Subscribe Model 12
2.2.3 QoS Parameters in ROS 2 DDS 12
2.2.4 Retransmission Mechanism of DDS 13
2.2.5 Differences Between DDS and TCP 13
2.2.6 Related Work: ROS 2 Performance Analysis 14
2.3 Industrial Wireless Sensor-Actuator Networks 16
2.3.1 IWSAN Architecture 16
2.3.2 Relationship Between IWSAN and Robotic CPS 16
2.3.3 Related Work: Network Design for Control Quality 16
2.4 Reinforcement Learning and Multi-Armed Bandit 17
2.4.1 Multi-Armed Bandit Framework 17
2.4.2 Why Multi-Armed Bandit for IWSAN Path Selection 18
2.4.3 Non-Stationary MAB in Wireless Networks 18
2.4.4 Related Work: Reinforcement Learning in Wireless Networks 18
2.5 Mixed-Criticality Systems 19
2.6 Structured Research Gap Analysis 20
2.6.1 Classification Framework 21
2.6.2 Quadrant A: No Analysis, No Adaptation 22
2.6.3 Quadrant B: Network Analysis Only 22
2.6.4 Quadrant C: Control-Aware Adaptation Only 23
2.6.5 Quadrant D: Both Analysis and Adaptation 23
2.6.6 Positioning of This Dissertation 23
2.6.7 Summary of Research Gaps 29
3 An Analytical Latency Model of the Data Distribution Service in ROS 2 31
3.1 Introduction 31
3.2 Motivation: ROS 2 DDS is different from TCP 32
3.3 Preliminaries on Data Distribution Service 34
3.3.1 Data-Centric Publish-Subscribe Model 34
3.3.2 Quality of Service Parameters in ROS 2 DDS 34
3.3.3 The Retransmission Mechanism of ROS 2 DDS 35
3.4 Latency Analysis of DDS Reliable Data Delivery 36
3.4.1 System Model 36
3.4.2 Derivation of the Steady-state Probability q 38
3.4.3 The Average Latency of Data Messages 41
3.5 Performance Evaluation 45
3.5.1 Testbed Configuration and Setup 45
3.5.2 Experimental Results and Observations 46
3.5.3 Analysis of the Average Latency 53
3.6 Related Work 54
3.7 Conclusions 55
4 CAPL: Criticality-Aware Adaptive Path Learning for Industrial Wireless Sensor-Actuator Networks 56
4.1 Introduction 56
4.2 Related work 58
4.3 Preliminaries on multi-armed bandit 60
4.4 System model 61
4.5 Criticality-Aware Adaptive Path Learning Algorithm
4.5.1 PathSelect Algorithm 64
4.5.2 OptDeltaSearch Algorithm 70
4.6 Performance Evaluation 72
4.6.1 IWSAN Framework 72
4.6.2 Simulation Results: Performance of CAPL 74
4.6.3 Experiment Results: Practicality of CAPL 76
4.7 Conclusions 78
5 From Parameter Analysis to Safe Environmental Search 80
5.1 Introduction 80
5.2 Network Parameter Controllability 81
5.2.1 Design Variables and Environment Variables 81
5.2.2 Formal Abstraction 82
5.3 Analysis: The Environment Variable as a Practical Bottleneck 83
5.3.1 Sensitivity to the Environment Variable p 83
5.3.2 Implications for Design Variable Optimization 84
5.4 Search: Finding the Best Environment Variable 84
5.4.1 The Search Problem and Its Cost 85
5.4.2 Control-State-Guided Search 85
5.4.3 The Effective Environment Variable 86
5.5 Complementarity of Analysis and Search 87
5.5.1 What Analysis Cannot Do 87
5.5.2 What Search Cannot Do 87
CONTENTS xi
5.5.3 The Criticality–Latency Interaction 87
5.6 Design Guidelines 88
5.6.1 Guideline 1: Design Variable Optimization via the Analytical Model 88
5.6.2 Guideline 2: Environment Variable Search via Control-Aware Exploration 89
5.6.3 Guideline 3: Joint Optimization 89
5.7 Generalizability and Limitations 89
5.7.1 Protocol-Agnostic Scope 89
5.7.2 The Value of Cross-Stack Validation 90
5.7.3 Limitations 91
5.8 Summary 91
6 Conclusion and Future Work 93
6.1 Summary of Contributions 93
6.2 Limitations 95
6.3 Future Research Directions 95
6.3.1 Extensions of the Latency Model 96
6.3.2 Extensions of the CAPL Algorithm 96
6.3.3 Cross-Layer and System-Level Directions 97
6.3.4 Broader Application Domains 98
국문초록 117
URI
https://scholar.dgist.ac.kr/handle/20.500.11750/60749
http://dgist.dcollection.net/common/orgView/200001011780
DOI
10.22677/THESIS.200001011780
Degree
Doctor
Department
Department of Electrical Engineering and Computer Science
Publisher
DGIST
Show Full 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???: