Cyber-Physical AI
Research Program, Sanghoon Lee (DGIST)
← Back to homepage
What makes intelligence and the physical world one system? Research on how AI, middleware, and control are designed together so that intelligence reaches the physical world reliably, despite real-world uncertainty and resource limits.
Three Pillars
- The Brain, AI. Good things happen often. (Right most of the time.)
- The Nerve, ROS 2. All things stay connected. (Decisions arrive, uninterrupted.)
- The Body, CPS. Bad things never happen. (Nothing bad must ever happen.)
AI is only often right. But the physical system it enters demands to be always connected and to never fail. This research builds what bridges the two.
Research Architecture
CPAI is the research program's paradigm, and the Middleware Harness implements it as the execution layer. Within this structure, the three axes (Embodied Intelligence, ROS 2 Middleware, Networked Control) are interconnected. CPAI implements the paradigm as an execution layer (Harness). The Harness mediates and delivers the outputs of intelligence (Embodied Intelligence), observes middleware state and tracks the conditions of execution (ROS 2 Middleware), and guarantees control safety and isolation (Networked Control). The three axes share wireless field measurements, system context and assumptions, and task goals and resource constraints.
Axis 1, CPAI (The Paradigm)
Cyber-Physical Artificial Intelligence. Defining how far AI can be trusted inside a physical system.
Core question. As AI grows more capable, whether it can be safely trusted inside a real physical system raises an entirely different question. In an era where AI directly produces a robot's actions, what do we entrust to AI, how far, and under what conditions does that judgment stay safely governed?
Frame. CPAI is proposed as the paradigm that unifies CPS and Physical AI. Rather than trusting an uncertain AI's judgment as-is, it checks each output at the system boundary, bounds its resources and time, and reverts to a verified action when at risk. Determinism is reconstructed not by internal proof but by boundary enforcement and runtime assurance.
Role in the program. The governing paradigm. Defines the problem space and the conditions under which AI can be trusted in a physical system. Every other axis operates within this framework.
Publications.
- Cyber-Physical Artificial Intelligence: A Unified Framework for Embodied Intelligence. Sanghoon Lee (Advisor: Kyung-Joon Park). DGIST, Ph.D. Dissertation, 2026.
- Cyber-Physical AI: Unifying Cyber-Physical Systems and Physical Artificial Intelligence. Sanghoon Lee, Guoliang Xing, Xue Liu, Karl H. Johansson, Kyung-Joon Park. ACM TCPS (20th Anniversary), Invited, Under Review.
- From Hype to Reality: A Survey of Physical AI in Industrial Robotics. Sanghoon Lee*, Jiyeong Chae*, Jinhong Park, Kyung-Joon Park. Elsevier EAAI, Under Review.
- Cyber-Physical AI: Systematic Research Domain for Integrating AI and Cyber-Physical Systems. Sanghoon Lee, Jiyeong Chae, Kyung-Joon Park. ACM TCPS, Published 2025.
- Integrating ROS 2 and Physical AI: Architecture and Challenges. Sanghoon Lee, Jiyeong Chae, Kyung-Joon Park. ICUFN 2025, Published 2025.
- A Survey and Perspective on Industrial Cyber-Physical Systems (ICPS): From ICPS to AI-augmented ICPS. Jiyeong Chae*, Sanghoon Lee*, Junhyung Jang, Seohyung Hong, Kyung-Joon Park. IEEE TICPS, Published 2023.
Axis 2, Middleware Harness (The Execution Layer)
Implementing CPAI in real robot middleware, connecting the three axes.
Core question. The realization of CPAI must take place in middleware that can observe and govern both the AI model and the physical system. The goal is not to rebuild safety mechanisms from scratch each time, but to engineer them once into the middleware so that every robot system can reuse them.
Frame. Middleware Harness is the layer that implements the boundary enforcement proposed by CPAI in real robot middleware. Just as a harness turns a raw model into a governed agent, the Middleware Harness turns robot middleware into the enforcement layer of CPAI. It is where CPAI and the three axes (Embodied Intelligence, ROS 2 Middleware, Networked Control) converge.
3-Layer Structure (Connect, Observe, Force).
- Connect. Links the rules declared at the intelligence level to the middleware. Realizes PIT (Projection, Isolation, Transfer) as a ROS 2 profile. (Tied to Embodied Intelligence.)
- Observe. Observes the middleware state to bridge Connect and Force. Minimizes observation cost while maximizing information. (Tied to ROS 2 Middleware.)
- Force. Based on what Observe sees, enforces control, communication, and computation on the data plane. (Tied to Networked Control.)
Role in the program. The execution layer. Implements what CPAI proposes in real robot middleware, connecting the three axes without modifying existing systems.
Publications.
- Harness Engineering for Physical AI: Robot Middleware Is the Harness Layer. Sanghoon Lee, Jiyeong Chae, Kyung-Joon Park. Middleware 2026, Under Review.
- ros2probe: Non-intrusive, Kernel-selective Observability for Robot Operating System 2 Middleware. Jisang Yu*, Sanghoon Lee*, Kyung-Joon Park. Middleware 2026, Under Review.
Axis 3, Embodied Intelligence (Brain, AI)
How does an embodied intelligence perceive, plan, and act in the real world?
Core question. What does it take for a robot to perceive, plan, and act in the real world? When many robots localize, navigate, and divide work together, how is that intelligence built and coordinated?
Frame. Each problem is approached at the layer where intelligence meets the system. Scheduling connects learning to the factory floor. MAPF connects planning to real robot motion. Localization and navigation connect perception to the physical environment.
Role in the program. The testbed. Grounds the full CPAI framework in real multi-robot systems, from scheduling and MAPF to localization and LLM-driven navigation.
Publications.
- LIMA: Local Intersection Marshaling Algorithm for Multi-Agent Path Finding in Large-Scale Industrial Environments. Sanghoon Lee*, Jisang Yu*, Geon-pyo Kim, Kyung-Joon Park. Elsevier RAS, Under Review.
- Natural-Language to Operational Policies: LLM-Driven Costmaps for Industrial AMRs. Jiyeong Chae, Hyunkyo Seo, Heonjae Lee, Sanghoon Lee, Kyung-Joon Park. Elsevier RAS, Major Revision.
- Robust Localization in Dynamic Indoor Environments Using Structure-Guided Point-wise LiDAR Scoring. Yeong-gi Hong*, Donghyung Lee, Sanghoon Lee, Jiyeong Chae, Kyung-Joon Park. IROS 2026, Under Review.
- Time-aware Costmap for Smoother and Less Disruptive AMR Navigation With ROS 2. Jiyeong Chae, Hyunkyo Seo, Sanghoon Lee, Kyung-Joon Park. IJCAS, Published 2025.
- From Issues to Routes: A Cooperative Costmap with Lifelong Learning for Multi-AMR Navigation. Jiyeong Chae*, Sanghoon Lee*, Hyunkyo Seo, Kyung-Joon Park. JIII, Published 2025.
- PINMAP: A Cost-efficient Algorithm for Glass Detection and Mapping Using Low-cost 2D LiDAR. Jiyeong Chae, Hyunkyo Seo, Sanghoon Lee, Kyung-Joon Park. IEEE TIM, Published 2025.
- Deep Reinforcement Learning-driven Scheduling in Multijob Serial Lines: A Case Study in Automotive Parts Assembly. Sanghoon Lee, Jinyoung Kim, Yongsoon Eun, Kyung-Joon Park. IEEE T-II, Published 2024.
- Learning-enabled Flexible Job-Shop Scheduling for Scalable Smart Manufacturing. Sihoon Moon, Sanghoon Lee, Kyung-Joon Park. JMS, Published 2024.
- False Alarm Prevention through Domain Knowledge-driven Machine Learning: Leakage Detection in Water Distribution Networks. Sanghoon Lee*, Jiyeong Chae*, Sihoon Moon, Kyung-Joon Park. IEEE Sensors Journal, Published 2024.
- Graph-based Reinforcement Learning for Flexible Job Shop Scheduling with Transportation Constraints. Sihoon Moon, Sanghoon Lee, Kyung-Joon Park. IECON 2023, Published 2023.
Axis 4, ROS 2 Middleware (Nerve, ROS 2): Space, Time, and State
Where does the communication layer slow down, drop, and drift?
Core question. When many robots move together wirelessly and in real time, where does the middleware linking them slow down, drop, and drift? Three demands collide: spatial connectivity, temporal predictability, and state continuity.
Frame. The three dimensions of Space (spatial abstraction), Time (temporal predictability), and State (state continuity) define the trade-off space of ROS 2 middleware, and each research thread targets a specific region within it.
Role in the program. The core middleware axis. The Space, Time, and State framework defines the vocabulary for the entire research program, and open-source tools extend it into the ROS 2 community.
Publications.
- The Three Dimensions of ROS 2 Middleware. Sanghoon Lee, Taehun Kim, Angelo Corsaro, Kyung-Joon Park. ACM CSUR, Under Review.
- Dependency Chain Analysis of ROS 2 DDS QoS Policies: From Lifecycle Tutorial to Static Verification (QoS Guard). Sanghoon Lee, Junha Kang, Kyung-Joon Park. IEEE IoT-J, Under Review.
- Optimizing ROS 2 Communication for Wireless Robotic Systems (DDS Optimizer). Sanghoon Lee, Taehun Kim, Yeonwoo Choi, Kyung-Joon Park. IEEE IoT-J, Under Review.
- Probabilistic Latency Analysis of the Data Distribution Service in ROS 2. Sanghoon Lee, Hyung-Seok Park, Jiyeong Chae, Kyung-Joon Park. IEEE TVT, Under Review.
- Characterizing DDS Discovery Bursts in ROS 2 Multi-Robot Systems. Taehyun Kim, Sanghoon Lee, Kyung-Joon Park. ICTC 2026, Under Review.
- ROS 2 System Handbook. Sanghoon Lee. Book, In Preparation.
- Analytical Modeling of Discovery Storms in DDS and Zenoh for Large-Scale Robotic Networks. Yeonwoo Choi*, Sanghoon Lee*, Kyung-Joon Park. INFOCOM 2027, In Preparation.
- Discovery Storm: Scalability Analysis of DDS and Zenoh in Large-Scale Wireless Robotic Networks. Sanghoon Lee, Yeonwoo Choi, Jiyeong Chae, Kyung-Joon Park. INFOCOM NetRobiCS 2026, Published 2026.
- ROS 2 Middleware Survival Guide: Navigating Space, Time, and State. Sanghoon Lee, Kyung-Joon Park. ROSCon Global 2026 (Toronto), Accepted 2026.
- How to Send Large Data in ROS 2. Sanghoon Lee, Taehun Kim, Jiyeong Chae, Kyung-Joon Park. ICNP 2025, Published 2025.
- An Analytical Latency Model of the Data Distribution Service in ROS 2. Hyung-Seok Park, Sanghoon Lee, Kyung-Joon Park. INFOCOM 2025, Published 2025.
Axis 5, Networked Control (Body, CPS)
Reliability, Resilience, and Security of control over networks.
Core question. When a control loop closes over a communication network, how does it withstand delay, loss, and attack while staying safe? Where along the path from intelligence to the physical system does failure strike, and how does the system recover?
Frame. This axis covers reliability, resilience, and security when a control loop runs over a wired or wireless network, spanning controller reconfiguration, resource allocation and co-design, and attack-surface analysis and defense.
Role in the program. The axis that enforces the CPS guarantee. From controller reconfiguration and DoS to cross-layer attacks and 5G resource allocation, it ensures that nothing bad ever happens when control runs over a network.
Publications.
- Exploiting ROS 2's Blind Spot: Cross-Layer Attacks from Discovery to ARP Spoofing. Hyunho Ryu*, Sanghoon Lee*, Young-Sik Kim, Dan Dongseong Kim, Kyung-Joon Park. IEEE IoT-J, Under Review.
- DIDoS: Disturbance-induced Denial-of-service Attack in Networked Cyber-physical Systems. Sangjun Kim, Sanghoon Lee, Kyung-Joon Park. IJCAS, Published 2025.
- Vulnerability Analysis of Cross-Layer Attacks on ROS 2. Hyunho Ryu, Sanghoon Lee, Kyung-Joon Park. ICTC 2025, Published 2025.
- Priority-Driven Resource Allocation with Reuse for Platooning in 5G Vehicular Network. Tae-Woo Kim, Sanghoon Lee, Dong-Hyung Lee, Kyung-Joon Park. Sustainability, Published 2025.
- Learning-Enabled Network-Control Co-design for Energy-Efficient Industrial Internet of Things. Sihoon Moon, Sanghoon Lee, Wonhong Jeon, Kyung-Joon Park. IEEE TNSM, Published 2024.
- Real-Time Controller Reconfiguration for Delay-Resilient Cyber-Physical Systems. Sangjun Kim, Sanghoon Lee, Kyung-Joon Park. IEEE Access, Published 2022.
- Resilient Controller Reconfiguration under Network Delay Attack in Cyber-Physical Systems. Sangjun Kim, Sanghoon Lee, Kyung-Joon Park. ICEIC 2022, Published 2022.
This is a static text version of an interactive page. Enable JavaScript for the full interactive research map. For the complete profile, see the homepage.