Conference Program

FRMCS will underpin the next generation of railway communications, enabling a far wider range of services than GSM-R and forming a core part of Europe's digital railway transition. This talk explores the cyber security challenges of that evolution, from identity, trust, and inter-domain protection to segmentation, zero-trust thinking, and lifecycle resilience in long-lived railway systems. It argues that secure FRMCS deployment requires an architectural and governance perspective, not just technical controls, if railway communications are to remain resilient, interoperable, and trustworthy over time.

This talk addresses the challenge of implementing the new European law in critical embedded equipment and the different strategies that can be implemented to address it.

The Datacom Industry Association (DIA) has been founded to strengthen advocacy for continued investments in European R&D on critical Datacom technologies for the Internet in general as well as in a number of vertical industries. This presentation will highlight the efforts of DIA members in positioning the importance of protocol and communication research for the upcoming Framework Programme 10 of the European Union in a roadmapping exercise to develop what is called a "Strategic Research and Innovation Agenda". Dr. Dirk Trossen will outline not only what is being advocated but also how an organisation like the DIA can galvanise its members' expertise to ensure proper representation of key topics for the future European Datacom R&D community.

The rapid growth and penetration of AI and machine learning into critical embedded systems are reshaping the design of next-generation real-time and safety-critical systems. Across domains such as automotive, avionics, industrial automation, and intelligent transportation, AI is transforming the design, development, validation, and evolution of real-time systems, as well as their functionalities, through learning-enabled perception, decision-making, and adaptation. To respond to these shifts and help define a research agenda, together with colleagues from the real-time systems community, we have initiated a coordinated effort—through a series of workshops—to identify key challenges and opportunities at the intersection of real-time systems and machine learning.
In this keynote, I take a step toward forming a roadmap for learning-enabled and learning-assisted real-time systems by bringing together insights and perspectives from both research and industry. First, I present research directions distilled from this community effort, highlighting challenges that remain to be addressed and future research avenues proposed by academia. I then complement this perspective with observations from industrial practice and needs, including integration complexity and validation and certification concerns. By jointly examining these perspectives, I aim to identify key gaps and opportunities that can guide and ground the development of such a roadmap.

This keynote lecture traces the engineering path from dynamic positioning (DP) to fully autonomous maritime navigation. Dynamic positioning is presented as a mature, certified template: a hard real-time control loop that holds a vessel on station using only active thrusters, structured around estimation, control, and thrust allocation. The talk then examines Maritime Autonomous Surface Ships (MASS), whose first international framework, the IMO MASS Code, entered into effect on 1 July 2026. Building on this foundation, it considers where artificial intelligence enters the loop, in perception, decision making, and learned control, and why the assurance of such components remains an open problem. A structured comparison of DP and MASS follows, addressing redundancy, the shift from fail-safe to fail-operational behaviour, remote operation centres, communication layering, operator competencies, cybersecurity, and the implications for ports. The presentation closes with open questions on certification and education, arguing that the embedded systems community is well-positioned to address the safety, real-time, and verification challenges that autonomous shipping now poses.

Artificial Intelligence increasingly relies on large volumes of distributed data that are often sensitive, regulated, and difficult to centralize. In healthcare, Federated Learning (FL) has emerged as a promising paradigm enabling multiple institutions to collaboratively train AI models while keeping patient data locally under their control. However, real-world deployments have revealed important challenges related to privacy, cybersecurity, trust, model robustness, and regulatory compliance.
This presentation will provide an overview of recent advances in Secure Federated Learning for medical imaging applications, including privacy-preserving techniques, Fully Homomorphic Encryption (FHE), secure aggregation, trusted execution environments, differential privacy, and federated model governance. Drawing from practical experiences in collaborative healthcare AI projects, we will discuss the benefits and limitations of current approaches and highlight key lessons learned from deploying FL in highly regulated environments.
Building on these insights, the presentation will explore how Secure Federated Learning can be transposed to the transport and mobility sector. As connected vehicles, smart infrastructures, and mobility platforms generate massive distributed datasets, similar challenges arise regarding data ownership, privacy protection, cybersecurity, and cross-organizational collaboration. We will examine potential use cases, architectural considerations, and emerging research opportunities where privacy-preserving collaborative AI could accelerate innovation while maintaining trust and regulatory compliance.

This contribution presents an automated and tailor-made fault detection system developed for the quality inspection of valve bellows in an automotive production environment. The system, already deployed in production, combines complementary sensing technologies, including in-line profilometers, line-scan cameras, matrix cameras, and dedicated lighting configurations, to detect both geometric and visual defects at production speed. Profilometry data are processed using a classification model for welding defect detection, while visual defects are identified through image-based inspection modules relying on segmentation models. The complete solution integrates data acquisition, signal and image post-processing, AI-based defect diagnosis, PLC communication, and an operator-oriented HMI. By replacing manual visual inspection with consistent, traceable, and real-time automated decisions, the system improves quality control reliability and supports a zero-defect manufacturing approach for critical valve bellow components used in automotive applications.

The increasing adoption of Physical AI in cyber-physical systems is driving the deployment of advanced AI capabilities directly on embedded edge platforms, where real-time constraints, safety requirements, and cybersecurity concerns must coexist. This talk discusses architectural and system-software technologies enabling the secure execution of AI workloads on heterogeneous embedded platforms operating in critical domains such as transportation, robotics, industrial automation, and aerospace.
The presentation will introduce hypervisor-based virtualization and isolation mechanisms that allow multiple applications with different safety and security requirements to share the same hardware platform while maintaining strong separation and predictable behavior. It will then discuss approaches to protect AI-enabled systems against cyber attacks, including secure execution environments, runtime monitoring, and mechanisms to safeguard intellectual property by preventing AI model theft and unauthorized access to sensitive data.
Finally, the talk will explore the concept of multi-enclave AI execution, where multiple AI models with different criticality, safety, and security levels can coexist on the same chip within isolated trusted domains. This approach enables the consolidation of heterogeneous workloads while preserving security, resilience, and certification requirements, paving the way toward trustworthy Physical AI at the edge.

This presentation will focus on the results, observations, and quality assessment approaches derived from two of Multitel's projects in the smart transport domain:

Navauwal: concerning the remote control of inland waterway barges

Lusta-5G: works on the remote control of cranes in an inland waterway port

Network infrastructures have undergone a significant transformation in recent years, evolving from human-centric and manually operated environments to autonomous software-defined multi-domain management systems. This evolution towards autonomous networking has further accelerated with the advent of agentic AI, which has the potential to improve the efficiency and responsiveness of network service deployment, monitoring, and management. Nonetheless, the adoption of agentic AI in network management also raises important concerns related to security, reliability, and resilience. In this talk, we will explore current architectural trends in autonomous network management with agentic AI and discuss the opportunities and challenges associated with delivering network services for critical embedded systems, with a particular focus on resilience. Finally, we will discuss architectural principles and mechanisms for building resilient and trustworthy autonomous network services.

Artificial intelligence is rapidly transforming robotics and embedded systems, enabling applications ranging from autonomous transportation and multi-robot coordination to intelligent manufacturing. At the same time, data-driven learning raises important challenges related to reliability, safety, scalability, and resource efficiency. In this talk, we discuss how modern control theory and optimization can provide principled tools to address these challenges. The presentation focuses on recent advances at the intersection of control systems, machine learning, and distributed optimization, with emphasis on learning-enabled robotic and cyber-physical systems. Topics include model-free coordination algorithms for multi-agent robotic networks, resource-efficient learning methods for embedded systems, and optimization-based approaches for learning in dynamic environments. We discuss how concepts from feedback control and output regulation can be used to design learning algorithms with guarantees on stability, convergence, and robustness. Applications to robotic coordination, autonomous systems, and transportation-inspired networked systems will be presented. Overall, the talk highlights how control-theoretic methods can help bridge the gap between modern AI capabilities and the reliability requirements of real-world embedded systems.

Verification of system correctness and security is an essential part of system development, especially for critical systems. Formal methods provide mathematical guarantees for verification results, yet due to their computational complexity, they are rarely used in industry. Statistical Model Checking is a formal technique that avoids the unscalability issue and can operate on large and complex systems. This presentation shows our observations from the application of Statistical Model Checking to systems in different domains for safety, security, and robustness properties.