Ultra-wideband (UWB) localization is one of the most promising indoor localization methods. Yet, non-line-ofsight (NLOS) positioning scenarios remain a challenge and can potentially cause significant localization errors. In this work, we leverage the utility of a group of mobile robots to test and validate our approach systematically in a real world setup. We use a particle filter based localization algorithm, which is wellsuited for accommodating arbitrary observation models, with the ultimate purpose of integrating various sensory information within a single framework. In particular, we propose a novel, probabilistic UWB TDOA error model which explicitly takes into account NLOS, and introduce it into our localization framework in combination with a standard motion model based on deadreckoning information. We subsequently extend our single-robot localization framework to a multi-robot, collaborative system by enabling the sharing of relative, inter-robot observations. Our experimental results show how the novel TDOA error model is able to improve localization performance when knowledge of the LOS/NLOS path condition is available. These results are complemented by additional experiments which show how a collaborative team of robots is able to significantly improve localization performance when poor knowledge of LOS/NLOS path condition is available.