For a longer motivation, please see this post.

Multi-agent control has historically been dominated by central control where each agent is sent a command from an all-knowing central station. The practicality of this approach is falling off a cliff as more practitioners are interested in deploying a) real world systems (where ideal communications is a challenging research problem) and b) high agent-count systems (often called swarms) where scaling laws come into effect. Centralized control breaks down simply because it is hard to get all the commands through the network and computation load increases dramatically.

The common solution is decentralized control. In centralized control, agents decide how to act based on signal from their neighbors. In reality, the concept of neighbors of a networked robot is not as stable as the concept of distance between the two. Communication connections drop and re-emerge, and the control systems of the future must be able to accomodate this.

In this paper, we released a new tool for studying the effects of and building new tools to empower decentralized or partially-centralized control in high-agent count settings. The paper describes these problems and future directions in great detail!