A Bayesian network framework based on Localization Risk Maps for anticipating and minimizing localization errors

Adequate localization in mobile robotics is essential for any task to be accomplished effectively and safely, particularly in changing environments. Modern localization systems are generally reliable, yet their accuracy remains conditional on scene structure and environmental stability. In feature-poor or dynamically changing environments, failures still occur and can critically impact downstream tasks. We present a framework to predict localization risk from environmental and task cues, enabling systems to anticipate and mitigate failure. For that, we introduce a map-based representation, a so-called Localization Risk Map (LRM), that spatially quantifies the level of available features as well as the level of dynamic changes by relying on state-of-the-art measures; after statistically validating them as suitable predictors of localization error, we define a Bayesian network-based model and perform parameter learning on it. By relying on a LRM and other variables, this model is able to estimate the level of localization error that might be obtained along a proposed path. We validate our method through simulated and real-world experiments with different mobile robots. The results demonstrate that our proposal consistently achieves a reliable prediction of localization error, taking into account the distribution of geometric features, the expected level of dynamic changes, or both. Also, we deploy our method in a motion planning use case, showing that the localization risk maps aid in finding the paths with the lowest localization error.