Multi-robot localization is a pivotal challenge for autonomous service tasks, providing teams of terrestrial and aerial platforms with the precision and coordination necessary for complex mission execution in dynamic environments.

This paper presents a novel distributed multi-robot localization framework that integrates Ultra-Wideband (UWB) measurements with Gaussian Belief Propagation (GBP) on factor graphs. Addressing key challenges in noisy, GPS-denied environments, the framework employs a signal quality estimator based on novelty detection via an overcomplete autoencoder neural network trained under ideal Line-of-Sight conditions. The resulting novelty score is embedded into the factor graph optimization along with tailored non-linear robust factors for UWB range data, enhancing the resilience of the system to interference, multi-path effects, and non-line-of-sight conditions. Extensive evaluations in both simulated and real-world settings demonstrate a significant reduction in localization error achieving an approximate improvement of 40% with GBP converging at 40 Hz for networks of up to 100 robots. This work marks a significant step toward robust, scalable, and distributed localization for complex multi-robot systems.

The study further offers insights into integrating adaptive sensor quality assessment within decentralized probabilistic inference frameworks, paving the way for future advancements in multi-robot localization.