*3.6. Shared Knowledge*

We already mentioned that our viewpoint of robot teamwork is closely linked to mutual awareness in the robots, i.e., teammates have—in addition to their local knowledge—some knowledge about their colleagues in the team. This awareness often, but not necessarily, is based on the provision of a shared global knowledge base for the entire team. The knowledge base, which typically would be implemented as a distributed replicated knowledge store, contains the concepts, objects, and relations known to the robots, as well as the fused perceptions of the state of the execution environment. In many works, the individual local view of a robot is called the *local world model*, while the shared team knowledge base is called the *shared world model* [57].

Potentially, such a shared knowledge base with frequently consulted thousands of objects and relations is a very resource-consuming and a performance-critical element of the teamwork. This creates several challenges, as listed below.


While there is a large set of publications focusing on knowledge representation techniques for robotic applications (e.g., [58–60]), little was published specifically on distributed knowledge bases for multi-robot systems. One thread of research—in particular for service robots—looked at offloading the knowledge base to the cloud [61]. The viability of such an approach clearly depends on the application requirements in terms of access performance and availability. Other approaches exploit a decentralized storage of knowledge [57] leading to well-known questions of consistency in data replication.

The spectrum of diverse requirements in multi-robot applications seems to be so large, that a harmonization of techniques for knowledge representation and storage is out of reach. On the other hand, a vast body of know-how is available for knowledge representation, reasoning, distributed data

storage systems, and replication and consensus protocols. In this situation, it is worth considering the question of whether the variety of existing solutions could be narrowed down to a few solution patterns, which would satisfy the majority of application scenarios.
