There are tasks that cannot be done by a single robot alone. A group of robots collaborating on a task has the potential of being highly efficient, flexible and robust. If one robot fails, another robot could take its position. However, coordinating a large group of robots through a centralized control unit is difficult as it would require the centralized unit to talk to all robots and compute next actions for a possibly huge number of team members. We investigate decentralized control strategies, where each robot is a self-contained unit able to communicate with or observe its closest neighbor robots and make decisions based on its own observations. The goal is that such a team of self-contained robot units is capable to achieving a joint goal. Such a decentralized approach scales to robot teams of any size. Our research in this area particularly focuses on:
- Decentralized learning strategies that enable a team of robots to improve over time and
- Learning approaches that help us find decentralized control strategies for complex problems that we know how to solve in a centralized way but finding a decentralized strategy from intuition turns out to be difficult.