Two popular multi-agent teamwork structures, joint intentions [7] and shared plans [13], consider a team to be a group of agents that negotiate and/or contract with each other in order to initiate a team plan. The team forms dynamically and stays in close communication until the execution of the plan is completed. In contrast, the teamwork structure presented in this article supports a persistent team effort towards a common high-level goal in the face of limited communication.
The concept of the locker-room agreement facilitates coordination with little or no communication. Although it has been claimed that pre-determined team actions are not flexible or robust to failure [39], the locker-room agreement provides a mechanism for pre-defining team actions with enough flexibility to succeed. In particular, set-plays are pre-determined team actions that can be executed without the need to negotiate or use extensive inter-agent communication: the locker-room agreement provides enough flexibility that the agents are able to seamlessly assume the appropriate roles.
While we use the term ``formation'' to refer to the largest unit of the teamwork structure, soccer formations are not to be confused with military-type formations in which agents must stay in precise relative positions. Despite this dual usage of the term, we use it because formation is a standard term within the soccer domain [18]. For an example of a multi-agent system designed for military formations, see [3].
Castelfranchi classifies different types of commitments in multi-agent environments [6]. In this context, locker-room agreements can be viewed as C-commitments, or commitments by team members to do the appropriate thing at the right time, as opposed to S-commitments with which agents adopt each other's goals. In the context of [8], the creation of a locker-room agreement is norm acceptance while its use is norm compliance. Within the framework presented in [24], the architecture is for interactive software and hardware multi-agents.
As mentioned in Section 3, the concept of behavior in the context of our team member agent architecture is consistent with that laid out by Mataric [22]. There, ``behavior'' is defined as ``a control law with a particular goal, such as wall-following or collision avoidance.'' Behaviors can be nested at different levels with selection among lower-level behaviors consisting of a higher-level behavior. Similarly, internal and external behaviors in our system maintain team coordination goals, physical positioning goals, communication goals, and environmental information goals (such as knowledge of where the ball is). These behaviors are combined into top-level internal and external behaviors.