A Distributed Approach for Fault Detection in Swarms of Robots

Swarm Robotic Systems (SRSs) are multi-robot systems usually composed of relatively simple robots. Local decisions and communication between robots allow for the emergence of complex behaviors of the entire SRS. The distributed nature of the SRSs enables their use in many real-world applications. However, despite the common belief that these systems are inherently robust and fault tolerant, it has been shown that even a few faulty robots could considerably hinder the performance of the entire SRS. In this paper, we propose a distributed fault detection approach that exploits machine learning classifiers to allow each robot of a SRS to detect faults in other robots and/or in itself. The proposed fault detection approach is data-driven, requiring a reduced amount of explicit domain knowledge, and is based on data that can be easily collected by common swarm robotics platforms. We test the proposed fault detection approach in simulation and analyze the results using non-parametric statistical tests. Our extensive experimental campaign shows that our approach has good performance and is robust regardless of the ratio of faulty robots in the SRS.