Effective interactions between the components of a multi-agent system are essential in maximizing the overall performance. We have studied effective communication strategies in terms of how to maintain connectivity and what information to communicate. The long-term goal is to develop principled methods to design, analyze, and control interconnected groups of heterogeneous agents.
Connectivity Maintenance
One aspect of communication is the connectivity maintenance, where the question being asked is “where should the agents move?” The connectivity issue becomes important when the robots have range-limited communication, and at the same time they are performing tasks that requires them to be spread out in the environment, for example, surveillance and monitoring. We have proposed a dynamic and modular formation design, that ensures intermittent communication while being spatially distributed in the environment.
- Modular robot formation and routing for resilient consensus: C-8 [ACC20]
Communication Policies
We have also studied communication strategy, where we ask the question “what should the agents say to each other?” and “how should they react to the received messages?”
In general, synthesizing a decentralized algorithm is challenging because we need to identify (i) the information required by each agent, and (ii) how each agent will use this information. We designed a training methodology using neural networks to learn task-oriented communication semantics using the example provided by a communication-unaware/centralized expert policy [C6]. As a case study, we successfully trained a decentralized perimeter defense strategy by imitating the centralized one designed in [C5]. This work will help us build a generic method to decentralize a known centralized policy.
- Decentralization of multiagent policies by learning what to communicate: C-6 [ICRA19]