Multi-robot localization: gaussian belief propagation on factor graph

The development of multi-robot swarm technology represents a transformative frontier in robotics, offering vast potential across industries such as disaster response, environmental monitoring, logistics, and large-scale agricultural operations. In this scenario, the ability of robots to effectively localize themselves becomes paramount. However, traditional centralized methods quickly become impractical, suffering from communication bottlenecks, limited scalability, and vulnerability to system failures. This challenge underscores the need for distributed solutions, where each robot independently contributes to the overall system, ensuring robust and scalable performance. The aim of this thesis is to investigate the capability of factor graphs to model the complexities of multi-robot systems and examine how Gaussian Belief Propagation (GBP) enables efficient inference on such models. Specifically, this research focuses on applying these methods to the problem of multi-robot localization, achieved through the fusion of odometry and ultra-wideband (UWB) measurements in a decentralized robotic swarm. Hence, this work aims to provide a robust, scalable solution for accurate localization in large-scale autonomous systems. The performance of the proposed solution has been evaluated through both simulations and real-world experiments. An ablation study is performed in simulation to study the scalability of the solution increasing the number of robots and the noise. Additionally, real-world testing will be conducted using a swarm of four TurtleBot3 robots, with the Vicon motion capture system serving as the ground truth reference. The system effectively integrates asynchronous data from ultra-wideband (UWB) signals, odometry, and inter-robot positional information, achieving a global localization error of under 15 cm in real-world experiments. The solution has proven scalable from single-robot setups to fleets of up to one hundred.