Isolating sub-populations to exploit locality in discounted robot foraging

We examine a canonical multi-robot foraging task, in which multiple objects must be located, collected and delivered. Each type of object must go to a unique delivery location. The value of each delivery is discounted over time. We describe a system in which a population of robots are effectively allocated to local (and thus high-reward-rate) foraging tasks, by keeping them ignorant of distant (thus poor-reward-rate) tasks. Robots learn about available tasks by local communication, with a fixed communication range that controls the rate at which task knowledge propagates. Our empirical data suggests that there is an optimal communication radius for our setting. Our system is effective at allocating robots to tasks, performing better than fully-informed robots. Interesting emergent group behaviour dynamics are described.