Control optimization of an aerial robotic swarm in a search task and its adaptation to different scenarios

Abstract In many cases, the control system for robotic swarms are complex behavioral networks (frequently probabilistic finite state machines) with several parameters to be tuned. Their selection has a high impact on the performance of the swarm carrying out a given task. In the past, those parameters have been optimized using automatic methods, whereas in other cases the network is simplified and tuned making use of expert knowledge. The problem becomes trickier when the performance depends not only on these control parameters, but also on other variables that cannot be selected by the designer (such as the size of the scenario, or the number of available agents). Moreover, there usually exist factors that inject noise in the measured outcome, such as the initial conditions, making the task more difficult to be analyzed. This work proposes and compares principled methods that address these two issues: the optimal configuration of controls with a high dimensional configuration space, that at the same time must be optimized for a broad range of scenarios. As a testbed task, search on a rectangular area is studied. We show that our proposal successfully addresses this complex problem. Moreover, our approach may be also implemented to configure subtasks inside a global mission, or on single behaviors that must be configured on-line depending on external state values.