Fault Recovery in Swarm Robotics Systems using Learning Algorithms

When faults occur in swarm robotic systems they can have a detrimental effect on collective behaviours, to the point that failed individuals may jeopardise the swarm's ability to complete its task. Although fault tolerance is a desirable property of swarm robotic systems, fault recovery mechanisms have not yet been thoroughly explored. Individual robots may suffer a variety of faults, which will affect collective behaviours in different ways, therefore a recovery process is required that can cope with many different failure scenarios. In this thesis, we propose a novel approach for fault recovery in robot swarms that uses Reinforcement Learning and Self-Organising Maps to select the most appropriate recovery strategy for any given scenario. The learning process is evaluated in both centralised and distributed settings. Additionally, we experimentally evaluate the performance of this approach in comparison to random selection of fault recovery strategies, using simulated collective phototaxis, aggregation and foraging tasks as case studies. Our results show that this machine learning approach outperforms random selection, and allows swarm robotic systems to recover from faults that would otherwise prevent the swarm from completing its mission. This work builds upon existing research in fault detection and diagnosis in robot swarms, with the aim of creating a fully fault-tolerant swarm capable of long-term autonomy.