*3.7. Robustness and Dynamic Adaptation*

Robustness is the ability of a system to cope with errors during execution and with erroneous input (https://en.wikipedia.org/wiki/Robustness\_(computer\_science)). A lack of robustness in dynamic team coordination may be due to various technical causes. Communication links may be unreliable, i.e., different communication technologies and conditions in the runtime environment may lead to message loss and network partitions such that standard communication protocols cannot provide a guaranteed error-free service. Individual robots may move out of communication range and be temporarily disconnected from the team. This has implications for the design of the application-level protocols. Centralized configurations are not appropriate in this case, since a single point of failure and performance bottleneck can severely hinder the teamwork and make the whole MRT useless. Sensors of team members may deliver different values for the same environment variable. This raises the question of what level of agreement the application requires for collective perceptions and whether the fusion of different types of sensor information can help in such a situation to improve the quality of the information in the shared world model. The amount of overhead for achieving reliable sensor information and consensus building may be prohibitive in very dynamic environments where swift decisions are more important than lengthy computation and communication activities. Moreover, run time execution errors may be caused by situations that were not foreseen at design time. Robustness in this case would mean that the system is able to perform an unanticipated adaptation. Thus, the team as a whole should be able to evolve its team plan, as well as the plans of the individual team members based on, e.g., input from other agents or machine learning techniques. In general, unanticipated software adaptations, which were not planned by the developer at design time, are a challenging problem. Only a few attempts on a general solution for unanticipated on-the-fly adaptation appeared in the literature [62,63]. Most adaptive systems assume that the adaptation state space is known completely at design time [64].

Related research questions are as follows:


In teamwork scenarios, the arrival of new team members with new capabilities or the departure of team members with individual capabilities might require changes in the team plans. Likewise, the evolution of global team goals and/or individual robot goals might demand a re-planning. Note that we are not concerned about the generation of the new plans. This may be done manually by a human developer or automatically by machine learning techniques and planning algorithms. Our emphasis is on the implications of openness of teams and, thus, on the capability for dynamic evolution and interchange of team plans.

A possible approach to unanticipated adaptation in an MRT is based on semantic annotations of team plans using a declarative logic programming language such as answer set programming (ASP) [65,66]. ASP adheres to a similar programming model as Prolog [67]. A number of projects showed that ASP meets the requirements for semantic specifications in a wide range of application areas in terms of expressiveness, efficiency, dynamic extensibility, and scalability. Examples are semantic service adaptation [68], dynamic information stream configuration in crisis management scenarios [57], and service robotics [59]. Thus, by adding semantic annotations to team plans, the developer lays the foundation for re-planning at runtime based on the specified properties and constraints for the

robots and their relationships. The semantic compatibility of annotated team plans can be checked using established techniques for semantic matching and adaptation [10,69]. Nevertheless, this is still largely unchartered territory in respect to the applicability to different robot scenarios. More research and practical experience are needed on the scope, expressiveness, and performance implications of different paradigms for unanticipated adaptation.
