Motivation and Purpose of the Seminar

Manufacturing cells and factories, transportation systems and various other parts of critical infrastructure such as energy grids have traditionally been controlled via self-contained, centralized systems continuously monitored and reconfigured by humans. The ever-growing complexity and integration of these Cyber-Physical Systems (CPS) into reconfigurable value chains (``Industrie 4.0''), autonomous cars and other services with high reliability requirements necessitates a radical change in the control strategy: Classic controllers will not be able to handle the massive amounts of data generated by these emerging systems, not only because of restrictions with regard to computational power and complexities that might bar human interventions in the processes, but also due to missing or inadequate methods for the control and the interconnection of the devices comprising such systems. Whilst CPS have moderate bandwidth/throughput requirements, often in the range of a few bytes per control or sensor message, they require high delivery success rates and predictable latency bounds for these messages and the computations performed on the data, often in the order of a few milliseconds. Stable controllers can only be developed if a predictable behavior of the communication and computation infrastructure may be assumed. Otherwise, the systems may not reach the desired states or even become unstable, up to the point where they may cause physical injuries or the loss of human life. Hence, a paradigm shift towards real-time oriented communication and computation in CPSs is necessary.

Such a shift can, however, only be achieved by overcoming the traditionally loose coupling in the design of system components in networks. Currently, both the communication systems community and the control systems community consider the components of the respective other field as a ``black box'' and abstract from the variations. Valuable insights that the other domain might provide towards the joint goal of keeping a CPS controllable may hence not be available. Although solutions have already been developed that bring communication and control closer together for specific use-cases, the abstraction problem has not been approached from a general, overarching perspective.

The purpose of this seminar was hence to bring together experts working in the key communities relevant for the science of CPSs and Cyber-Physical Networking (CPN) to get a clearer and more detailed picture of the most important issues of the control and networking aspects that CPSs/CPNs bear and to identify the mutual relations and influences of the associated fields, in order to overcome the so-far strict abstractions and boundaries that exist, and to sketch a roadmap for further research in the field. The driving question was how it is possible to derive generalizable co-design methods and metrics that support the development of universal networked CPSs/CPNs.

Prior Dagstuhl Seminars have already addressed CPS aspects such as synthesis (Seminar 17201) and verification methods (Seminar 14122), robustness (Seminar 16362), as well as software engineering for control (Seminar 14382), yet none of these have focused on the interaction, interdependencies and the co-design of communication and control.

Participants and Structure

The seminar brought together a total of 30 participants from various fields within the communication and control domains, ranging from promising young scientists to leading authorities within their respective fields, but also including practitioners from industry with a strong research background, as well as representatives from funding organizations.

The first day of the seminar was dedicated to an in-depth introductory session. Besides as short personal introduction with background and current research interests, each participant was asked to prepare a personal statement answering the following questions:

• What are the most important problems to solve in the realm of CPS/CPN?
• What are the main scientific challenges and which fields can contribute to them?
• What have we achieved so far, and what are the pitfalls of past and current research?

Each personal statement was followed by a discussion round on the presented individual statements. The statements and discussions proved highly fruitful, as they allowed the organizers and the participants to gain an understanding of the current state and future challenges in the Control of Networked CPS from the different disciplinary perspectives.

Most often, opinions revolved around the need to understand more about the implications of the dynamic behavior of both the controlled systems themselves and of the communication networks. Research so far seems to have primarily focused on the "steady state", as participant termed it. The uncertainties introduced by controlled systems and (especially wireless) networks in coexistence with other systems, however, seem to call for various improvements in CPS/CPN design. Yet, as other participants expressed it, besides having fostered a better understanding of the basics of the respective other fields in recent years by programs such as DFG's Priority Programme 1914 Cyber-Physical Networkin, "little" has been achieved be community so far, with a major pitfall being "lopsided" methods which are often attributed to "sticking to domain-specific models". Opening these models to incorporate knowledge from other domains, therefore, seems to be a major challenge for the upcoming time.

A further major topic discussed was the need for more realistic and relevant problem settings in the research efforts, since, as one participant put it, "real problems are more complex than a single inverted pendulum". Hence, to avoid "esoteric" research and thus "ending up as an academic field with zero practical impact", CPS/CPN is in the need of "prov[ing] that what we develop is useful/needed" within the upcoming years. This does not mean that basic research has or needs to be concluded in any way. Yet, further opinions voiced more than once regarded energy efficiency and usable abstraction/decomposition methods (which may at times even sacrifice optimality for applicability and efficiency) as interesting research challenges for the upcoming years, which shows that the community has already begun tackling more practical issues recently. A variety of additional comments showed that few, if any, of the issues of CPS/CPN can be considered as solved by today.

Plenary Discussion: Properties of Cyber-Physical Networks

The unexpected intensity of the discussions following the respective personal introductions revealed the extreme variety of opinions on the nature of CPS/CPN and the major challenges in this interdisciplinary field. To facilitate a common understanding, the personal introductions were thus followed by a plenary discussion on which properties define CPS/CPN and make them interesting for scientific study.

It was agreed that - besides the eponymous intertwining of control, networking and the physically tangible world - CPS/CPN are dominated by uncertainties of both the systems and their operational environments, dynamics of configuration and load, (usually) limitations e.g., with respect to the capacity of the network, computation power and energy, a control objective that is sought to achieve through the network (if it is not serving pure monitoring purposes), as well as the associated relative administrative and technical autonomy of CPN compared to their traditional counterparts. Regarding typical metrics of timing and scale, it was further agreed that traditional complexity metrics do not apply to CPN. There often exist intricate and counterintuitive relationships between timing constraints of control and the network, leading to situations in which certain upper- and lower(!)-bounded delays may even be beneficial for the simplification and stabilization of control. Hence, defining the time-criticality of a system is scenario-dependent. Likewise, scaling effects may lead to situations in which too many local observations may prove counterproductive to controllability so that, depending on the scenario at hand, issues arise regarding the "right" amount of information sharing between local and global players in distributed decision-making processes. As such, conceiving widely-applicable categories for the complexity of CPS/CPN was identified as an open problem.

Impulse Talks & Plenary Discussions

For the remaining one and a half days of the seminar, the participants were asked to propose impulse talks on topics related to their respective areas of control of CPS/CPN research. Each talk served as the basis for a subsequent plenary discussion aimed at identifying worthwhile research directions for the community. extbf{Out of a total of 18 proposed impulse talks, six talks were selected by the organizers.} In the following, we present the major insights from the talks and the discussions.

• The development of next-generation wireless communication technologies such as 5G and the increased efficiency of small-scale mobile devices in general, have fueled the interconnection of ever more devices into large-scale CPS. However, as the number of devices generating data and potentially taking action increase, so do the burdens on controllers and the network. In his talk, Carlos Canudas-de-Wit showed first results pointing at the fact that both state estimation and control may provide sufficient results even when only considering a well-chosen aggregating subset of a system's sensing and actuating nodes, as long as the distribution of these nodes follows a specific structure, which can, however, be found for many real-world scenarios. Together with another technique based on partial differential equations, the results of his work showed that when combining both control- and information-/network theoretic models, as well as upcoming techniques such as in-network computation that may provide the necessary aggregation infrastructure, even systems of immense complexity can be controlled without overloading controllers and networks.
• Another challenge of CPS arises when safety guarantees need to be fulfilled, especially when a failure to meet these guarantees can lead to injury or endangerment of human life. Adam Molin presented an industry perspective on the validation and verification of (increasingly) autonomous vehicles, a field in which scenario-based testing approaches represent the state-of-the-art. While the determinism of systems without humans in the loop may aid in the construction of such scenarios, only probabilistic guarantees can be given when humans are involved in the operation of a system. This fact reflects not just on automotives but on multiple other scenarios discussed in the seminar and highlights the importance of joint analysis methods for the control and the communication components of such systems.
• In her talk on 5G Service Automation, Chrysa Papagianni expressed the view that upcoming mobile networks will witness a shift from open-loop to hierarchical closed-loop control as customers shift towards a pay-per-use scheme for the offered services. Whether the control problems (e.g., regarding network slicing) can be considered to exhibit sub-minute or even real-time requirements (as witnessed in most other systems discussed in the seminar) is still an open question. Yet, considering the anticipated, wide-spread application scenarios of 5G also in the area of CPS/CPN, the seminar identified the issue of base station multi-tenancy as an area for future research within the context of CPN.
• A cornerstone for the successful operation of CPS/CPN are easily-calculable metrics to assess the operational status, as well as to guide the generation, transmission and evaluation of signals within the systems. Vahid Mamduhi in his talk showed that simple age-of-information (AoI) -- a common metric applied both by the control and the communications communities in theoretical and practical scenarios -- bases on assumptions that can hardly be met by the systems. As a consequence, AoI needs to be augmented by notions of state, timing constraints of the system, and the objective of the control function (all related to a single piece of information) to really provide benefits. Such metrics are arguably hard to conceive for the general case, yet the talk inspired discussions among the participants regarding sensible metrics with broader applicability.
• From a more communication-oriented perspective, James Gross presented his group's efforts towards determining latency bounds in wireless CPS/CPN. Both a queuing-theoretic and a model checking-based approach (the latter concerning a practical implementation of an ultra-reliable low latency protocol) yielded qualitative results that seem promising. Yet, the practical applicability of such approaches is currently hampered by assumptions regarding distortion that may not hold in practice. In the subsequent discussion, topics included (a)the question whether making the network completely deterministic (or the ability to make determinism assumptions) is actually needed and achievable, and how possible compromises may look like, b) to which degree techniques such as software-defined networking, in-network processing, time slicing and standards such as 5G can contribute towards such goals, and (c) which interfaces, abstractions and design patterns should exist that allow specifying and proving certain guarantees in CPN, especially regarding the interplay of control algorithms and networks.
• The complexity and variety of communication protocols within automation is addressed by the recent Time-Sensitive Networking (TSN) efforts of the IEEE, which seek to offer a vendor-neutral Ethernet-based solution catering both legacy and future real-time applications, including control. Eventually, the automation pyramid will be transformed into an automation pillar at which TSN serves as the (sole) connectivity provider for control loops which will span the whole automation network from virtualized (/centralized) controllers and the field level. In his talk, Tobias Heer provided an overview of the changes that TSN brings with regard to medium access methods to enable real-time capabilities in Ethernet. While TSN brings significant improvements to wired settings, the subsequent discussion round revolved around the difficulties in achieving this in wireless scenarios. Besides the apparent issues of jamming and/or other attack vectors in wireless control systems, the possibility of trading reliability against capacity and the resulting implications on control algorithms was identified as a research issue.

Conclusion

Throughout the presentations and especially the discussions both during the plenary sessions as well as during off-hour activities, the seminar successfully brought together researchers from control and communication from both academia and industry, and undoubtedly fostered a deeper understanding of the intricate interplay of the disciplines in the research area of CPS/CPN. A variety of open problems and promising research areas were identified, with some in dire need of increased cooperation between the involved fields. This underlines the need in CPS/CPN research for formats valuing open and honest discussions, and both the organizers and the participants hope to be able to continue these discussions in the following years through additional summits and - once the insights gained in this first edition have shown visible impact on the scientific community - possibly another Dagstuhl Seminar. As a concrete follow-up, the organizers and participant James Gross are planning to conduct a seminar in Stockholm/Sweden in 2020 on this diverse research area.

Creative Commons BY 3.0 Unported license

René Glebke, John S. Baras, Sandra Hirche, Kay Römer, and Klaus Wehrle