Certifiable, Resilient, and Efficient Real-Time CPS: A Journey in Embedded AI
When I first encountered an intelligent system missing a deadline—not because of a hardware fault, but due to a poorly scheduled inference—I knew something had to change. That realization sparked a journey that led to more than a decade of research, a vibrant lab at NC State, and ultimately, the honor of receiving the 2025 SIGBED Early Career Researcher Award.
In this post, I’ll share how that journey evolved—through research that sits at the intersection of real-time scheduling, machine learning, and embedded systems—and how I envision the future of intelligent cyber-physical systems (CPS) that are not only capable, but also certifiable, efficient, and resilient.
When Real-Time Systems Meet AI
CPS applications—from autonomous vehicles to wearable health devices—are becoming increasingly intelligent, but with this added intelligence comes added complexity. These systems must respond in real-time under uncertain conditions, with limited resources, and often in safety-critical scenarios. The traditional design principles for embedded systems struggle to scale when modern workloads include perception, learning, and control, all running concurrently.
To address this, my research focuses on three tightly connected levels: real-time scheduling and resource management at the system level, adaptive and lightweight inference at the algorithmic level, and trust and security at the data level.
Mixed-Criticality Scheduling: Building the Foundations
My early work focused on enhancing predictability in embedded platforms using mixed-criticality (MC) scheduling models. We designed precise scheduling approaches for varying-speed multiprocessors and host-centric architectures, enabling systems to dynamically adapt to critical workloads while maintaining temporal correctness. We extended these ideas to gang tasks, co-scheduling on CPU-GPU systems, and even to systems with hierarchical pacing and data offloading. One of our more recent ideas, IDK-cascades, explores how uncertain classification outcomes in ML pipelines can be deferred or resolved adaptively to minimize the expected time to successful decision-making.
These techniques collectively form a foundation for designing systems that can reason about timing, uncertainty, and resource constraints in a principled way.
From Theory to Systems: Embedded AI in Action
Real-world deployments helped push these ideas further. In the F1-Tenth autonomous racing platform, we built a real-time ROS 2 framework with physics-informed scheduling, preemptable executors, and digital-twin-based safety validation. Our scheduling analysis for ROS 2 with resource contention and latency constraints showed that it’s possible to achieve both timing guarantees and runtime adaptability—something long thought difficult in robotics.
In wearable healthcare, we tackled real-time cardiovascular disease detection using embedded deep learning models. Our work introduced attention-based CNN-LSTM architectures that balance inference latency, accuracy, and memory use on resource-limited platforms. These systems adapt model complexity based on a patient’s heart signals, enabling both real-time responsiveness and power efficiency.
We also applied deep learning to wearable kinematic sensing, using compact IMU sensors and multimodal networks to reconstruct lower-limb motion and joint forces in daily environments. This work is particularly relevant for monitoring gait abnormalities and rehabilitation, where lab-grade motion capture is infeasible.
Trustworthy AI for Embedded Systems
Another critical direction is ensuring trust and safety when ML models are deployed in embedded CPS. We examined the growing risks of backdoor attacks in pretrained models and proposed defenses using Fisher information to guide purification. We also developed domain adaptation techniques that allow systems to adapt securely to changing environments without needing source data—essential for mobile and privacy-sensitive CPS.
Furthermore, we investigated federated learning in real-time systems, enabling multiple edge devices to collaborate without centralized coordination while still meeting strict deadlines. This line of work has implications for collaborative robotics, smart infrastructure, and multi-party sensing systems.
Looking Ahead
Our lab continues to explore how AI-enabled systems can be made predictable, secure, and efficient. We’re investigating physics-informed control for real-time robots, real-time federated learning protocols for mobile CPS, and adaptive digital twins that enable continuous verification of complex systems. I’m excited to collaborate with others working on the same grand challenge: making intelligent embedded systems trustworthy by design.
I’m very honored to receive the SIGBED Early Career Award, and I remain committed to pushing the boundaries of what embedded and real-time systems can achieve.
Author: Zhishan Guo is a tenured Associate Professor at the Department of Computer Science of North Carolina State University, where he directs the Real-Time and Intelligent Systems lab and the Cyber-Physical Systems (CPS) research group. His research interests are in real-time scheduling theory and machine learning theory with applications to cyber-physical systems. He is a recipient of multiple highly competitive awards, such as Best Paper Awards, Outstanding Paper Awards, and Best Student Papers Awards in EMSOFT, RTAS, and RTSS, Outstanding Undergraduate Teaching Award from UNC-Chapel Hill, etc.
He has been committed to strengthening the embedded and CPS community, such as serving in the IEEE TCRTS executive committee since 2020, and as the Treasurer of ACM SIGBED since 2025, supporting events, student programs, and the continued health of our community’s technical and organizational initiatives. As a co-founder and lead of the CPS Research Focus Group at NC State, he helps bring together about 10 faculty members and more than a hundred students across multiple disciplines to tackle problems at the intersection of sensing, communication, AI, and computation. These roles reflect his belief that technical progress goes hand in hand with community building.
Disclaimer: Any views or opinions represented in this blog are personal, belong solely to the blog post authors, and do not represent those of ACM SIGBED or its parent organization, ACM.