Cyber-Physical Artificial Intelligence: A Unified Framework for Embodied Intelligence

Abstract

본 연구는 사이버-물리 인공지능의 개념을 확립하고, 체화 지능을 사이버-물리 시스템과 인공지능의 통합으로 재정의한다. 사이버-물리 인공지능은 가상 영역에 머무르던 지능을 물리 세계로 확장하여 인지 · 지각 · 구동이 물리 공간에서 유기적으로 결합되도록 하며, 그동안 인공지능 모델이라는 두뇌에 치우친 연구 관행으로 인해 간과되어 온 물리 시스템이라는 몸과 미들웨어라는 신경계의 제약을 균형 있게 다룬다. 이를 위해 본 연구는 제약–목표–접근의 세 차원에서 인공지능과 사이버-물리 시스템을 체계적으로 연결하고, 인공지능 모델 전 주기 최적화와 미들웨어 최적화를 아우르는 통합 최적화 프레임워크를 제안한다. 먼저, 체화 지능을 위한 인공지능과 사이버-물리 시스템 통합의 생애 주기를 설계–개발–배포 단계로 정의하고 아홉 가지 핵심 과제를 도출한다. 설계 단계에서는 엣지 환경의 계산 제약과 클래스 불균형을 해결하기 위해 도메인 지식 기반 특성 선택기와 새로운 데이터 증강 기법을 제안하여, 센서 네트워크 기반 누설 감지에서 오경보를 유의하게 감소시킨다. 개발 단계에서는 데이터 희소성과 지연 보상으로 인한 불확실성을 완화하기 위해 실 데이터 기반 디지털 트윈을 구축하고, 지연 효과를 정밀하게 연결하는 보상 큐 기반 강화 학습을 도입한다. 그 결과 자동차 부품 다공정 직렬 라인의 스케줄링 정책 신뢰도를 향상시켜 산업 자동화로의 실행 가능성을 입증한다. 배포 단계에서는 라벨 희소성과 데이터 드리프트에 대응하기 위해 자동 라벨링 기반 자가 지도 학습과 평생 학습을 결합하여, 다중 자율주행 로봇 기반 물류 환경에서 장애물 변화에도 지속 · 적응적 성능을 유지한다. 논문의 후반부는 사이버-물리 인공지능을 미들웨어 계층으로 확장하여, 사이버–물리 상호작용의 핵심 인프라인 로봇 운영체제 2의 데이터 분산 서비스를 대상으로 통신 생애 주기 분석과 종합적 최적화를 수행한다. 첫째, 데이터 분산 서비스의 전송 · 재전송 메커니즘을 이산 확률 상태로 모델링하여 임의 패킷 손실률에서 지연 분포와 평균 지연을 도출함으로써, 불확실한 무선 환경에서 실시간 물리 인공지능 운용의 이론적 기반을 제공한다. 둘째, 대용량 페이로드 전송 실패의 병목인 IP 단편화, 재전송 타이밍, 버퍼 혼잡을 규명하고, 사용자 설정 파일 기반 경량 파라미터 최적화로 처리량–신뢰성 균형을 달성하여 제한적 무선 자원에서도 강건한 멀티모달 전송을 보장한다. 셋째, 데이터 분산 서비스의 디스커버리 단계에 대한 구조적 취약점을 밝혀 로봇 운영체제의 보안 설정을 우회하는 신규 계층 간 공격 체인을 제시하고, QoS 이벤트와 ARP 무결성 검증을 이용한 경량 실시간 방어 기제를 설계 · 검증하여 미들웨어 보안을 강화한다. 종합하면, 본 연구는 사이버-물리 인공지능이라는 개념적 틀 아래 인공지능 모델과 미들웨어를 함께 최적화하여, 데이터 처리 효율, 학습 신뢰성, 평생 적응성을 모델 계층에서, 실시간성, 신뢰성, 탄력성과 보안을 미들웨어 계층에서 동시에 달성한다. 스마트시티, 제조, 물류 등 실제 산업 시스템에의 적용을 통해 제안 기법의 실효성을 입증하며, 사이버-물리 인공지능이 안전한 자율성과 지속 가능한 지능을 지향하는 차세대 체화지능의 산업적 현실화를 위한 토대를 제공한다.|This study establishes the concept of Cyber-Physical Artificial Intelligence (CPAI), a unified research paradigm that redefines Embodied Intelligence as the seamless integration of Artificial Intelligence (AI) within Cyber-Physical Systems (CPS). CPAI extends intelligence beyond virtual environments toward physical AI, where cognition, perception, and actuation coexist within the physical world. Current research on physical AI has focused excessively on the performance of AI models, the system’s “brain”. This focus comes at the expense of considering the critical limitations of the “body,” the physical system, and the “nervous system,” the middleware that serves as the intermediary layer between the physical and cyber system. This imbalance has created a significant disparity between theoretical potential and real-world reliability. CPAI aims to overcome this structural imbalance. By systematically connecting the brain (AI models), nervous system (middleware), and body (physical system), CPAI establishes a new foundation for Embodied Intelligence capable of operating reliably under uncertainty and resource constraints. It clarifies the implementation of Embodied Intelligence as the integration of CPS and AI across three key dimensions: constraints, objectives, and approaches. This positioning establishes CPAI as an interdisciplinary blueprint for safely and efficiently embedding intelligence within physical processes. Building on this theoretical foundation, this research proposes a unified optimization framework for Embodied Intelligence, comprising the lifecycle optimization of AI models and the optimization of middleware. This study defines the lifecycle of AI–CPS integration as comprising the design, development, and deployment phases, and identifies nine major challenges. Based on these, the AI model optimization framework is proposed to systematically address the key challenges across the entire lifecycle. In each phase, the optimization strategies are proposed considering the physical constraints and operational environments of each system, and are supported by their applications to real-world industrial cases. In the design phase, data processing efficiency is enhanced to address limited computing resources and severe class imbalance. A novel domain knowledge–based feature selector, TNFS, and a GAN-based data augmentation technique, GES, significantly reduce false alarms in IoT-based leak detection systems, demonstrating that robust and resource-efficient physical AI can be achieved even at the edge level. In the development phase, the reliability of the training process is strengthened to mitigate data scarcity and uncertain inference caused by delayed rewards. A digital twin is constructed from real process data to safely generate rare fault scenarios, and a reinforcement learning formulation based on a redesigned MDP is developed to precisely link actions with delayed outcomes. As a result, the reliability of scheduling policies in a multi-task serial production line for automotive parts assembly is significantly improved, demonstrating the potential to transform reinforcement learning policies into executable automation for industrial assembly systems. In the deployment phase, the persistence of system adaptation is enhanced to address the challenges of label scarcity and data drift. Through the synergy of automated labeling–based self-supervised learning and life-long learning mechanisms, the system achieves continuous and adaptive performance in multi-AMR logistics environments. This maintains efficient navigation for multiple AMRs under changing obstacle patterns, and demonstrates continuous physical AI beyond the boundaries of initial training. The latter half of the dissertation extends CPAI to the middleware level, performing optimizations on the Data Distribution Service (DDS) of Robot Operating System 2 (ROS 2), the core infrastructure where cyber and physical layers interact. A lifecycle-based analysis of DDS communication is conducted to identify intrinsic constraints and major challenges that lead to communication failures. Building on this analysis, a comprehensive middleware optimization framework is proposed to address the major challenges of AI–CPS integration. First, the real-time performance of the middleware is improved to address data loss. By modeling DDS transmission and retransmission mechanisms as discrete probabilistic states, this study quantifies the statistical distribution model of delays and derives latency under arbitrary packet loss rates. This provides a theoretical foundation for real-time physical AI operation in uncertain wireless environments. Second, the communication reliability of the middleware is improved to address network resource limitations. It identifies key bottlenecks in large-payload transmission in DDS as IP fragmentation, retransmission timing, and buffer congestion. Based on this, it introduces a lightweight, XML-based parameter optimization framework that balances throughput and reliability with minimal overhead. This ensures robust multi-modal data transmission for physical AI even in wireless environments with limited network resources. Lastly, middleware security is reinforced to mitigate adversarial attacks. Previously unrecognized vulnerabilities in the DDS discovery mechanism are revealed, along with the design of new cross-layer attacks that bypass existing security features. A practical and lightweight defense mechanism is then proposed to detect and block these critical threats in real time, thereby fundamentally advancing the cybersecurity of Physical AI at the middleware level. In summary, this study presents a unified framework that optimizes and integrates AI models and middleware through the conceptual paradigm of CPAI. It unifies data processing efficiency, learning reliability, and lifelong adaptability at the AI model level, while ensuring real-time, reliable, and resilient communication at the middleware level. By validating the proposed approaches across a wide range of real industrial systems, including smart cities, manufacturing, and logistics, this research demonstrates that, through CPAI, AI can evolve into a new generation of Embodied Intelligence that is fundamentally efficient, reliable, adaptive, and secure. Ultimately, CPAI lays the foundation for transforming AI from a theoretical concept with potential applicability in the physical world into an industrial reality that coexists within it, guiding Embodied Intelligence toward safe autonomy and sustainable intelligence.