A Data-Driven Method to Identify Fault Mitigation Strategies in Robot Swarms

As robot swarms are increasingly deployed in the real-world, making them safe will be critical to improving adoption and trust. A robot swarm is composed of many individual robots each susceptible to failure at any given time, which may decrease the performance of the swarm as a whole. The ability to mitigate critical faults is therefore necessary. The difficulty with designing an effective mitigation strategy lies in the complexity of the swarm as a system, where individual in-
teractions give rise to emergent behaviour. In this paper, we present a data-driven method to identify effective local actions available to faulty robots in the swarm. We make the assumption that robots are able to self-detect faults and that pre-coded actions are indeed available. An effective action should mitigate any negative impact of faults on overall swarm performance. We consider two intralogistics scenarios where the swarm must retrieve and deliver boxes. The first concerns single robot transport (one robot per box) and the second, collective transport (four robots per box). Our method is able to identify effective actions for particular fault types. We also consider the impact of actions across ratios of fault in the swarm. Interestingly, faults do not always benefit from
mitigations, with mitigations causing overall lower system performance for certain fault types.