*3.1. Dynamic Coalitions*

In open robot teams where the team members are not known a priori at design time, we need support for establishing a temporary team membership. Participants form temporary coalitions in order to solve a problem and achieve a common goal. A traffic intersection in autonomous driving scenarios is an example of a short-lived coalition with continuous team reconfiguration, while Industry 4.0 scenarios would likely imply a longer-lasting coalition. Only members of a coalition should be involved in the teamwork interactions, and agents outside of the coalition should not disturb it. This requires that all agents know about their membership. Moreover, it may require security measures to protect the interactions of a team. Key challenges are as follows:


Many communication paradigms achieve interaction among distributed components based on the identities of the components. Examples are the classical client/server model, the actor model [14], or named channels in channel-based binary communication [15]. On the other hand, broadcast communication [16] may not require identities depending on the capabilities of the underlying communication system, but loses the ability to address a selection of individual agents. However, in open teams in dynamic environments, the identity of robots may not be known at design time, if robots may join and leave a team at run-time. Thus, the concept of identity is not easy to establish and may even be irrelevant [17].

In such environments, we need different ways to determine team membership and to address team members. Note that a single central team manager that monitors and controls team membership is out of the question here because we need to avoid a single point of failure in environments where robots may move out of reach temporarily or break down completely.

One solution is based on attribute-based interaction. It is a variant of publish/subscribe communication, and it was proposed in [17,18] as a paradigm to address collectives of possibly anonymous agents. In attribute-based interaction, robots of a multi-robot system explicitly expose a set of attributes that are relevant for the application at hand. Interaction between robots is based on groupcast communication, whereby sending and receiving messages is determined by predicates over the specified attributes. A *send* command expresses the intention to deliver a message to all robots satisfying the *send predicate*. Likewise, a *receive* command signals willingness to receive messages from team members according to the specified *receive predicate*.

For example, in an Industry 4.0 scenario, one might ask for "components that need to be delivered within the next 15 min" or, in autonomous driving, one might want to address "vehicles that are capable of autonomous driving and are closer than 50 m to the intersection". Thus, attribute-based interaction is a more fine-grained content-based selection of possible receivers and senders. Potentially, it allows a more efficient filtering of messages by the distribution infrastructure and reduces the communication overhead. The drawback is the need for a powerful, rather heavyweight distribution infrastructure. Attribute-based interaction is integrated into the syntax of several programming languages, such as Erlang [19] and Google Go [20].

Other well-known protocols for open coalitions include the JXTA peer-to-peer protocols that target overlay-based communication of peers across public networks, where security issues such as firewall traversal are of utmost importance [21]. An application of JXTA for the control of robots was reported in [22]. However, the JXTA project is no longer officially supported. The FIPA recommendations (http:// www.fipa.org/) offer a conceptual framework for communication and management in multi-agent teams. Their focus is on intelligent (software) agents, not on mechatronic robots. The FIPA standardization is currently not active.
