Parallel Graph Inference for Decentralized and Asynchronous Multi-Robot Cooperative Localization Under Limited Connectivity

Abstract

Limited connectivity poses a critical challenge for multi-robot cooperative localization, as delayed or out-of-order measurements often cause conventional algorithms to misalign and diverge. Existing methods typically struggle to balance real-time constraints with the computational cost of retrospectively fusing latent data. To address this, a parallel graph inference framework is proposed for decentralized and asynchronous cooperative localization. Each robot constructs a local sparse factor graph to uniformly represent state variables and measurements. Crucially, a multi-thread architecture is introduced to decouple high-frequency real-time estimation from the latency-prone fusion of inter-robot measurements. This allows the system to maintain control-rate state feedback while asynchronously integrating historical relative measurements via Gaussian belief propagation. Extensive evaluations on simulated and real-world datasets validate the system across varying sensor configurations, scales, and connectivity conditions. The results demonstrate that the proposed method remains robust under significant communication delays, achieving the accuracy of optimization-based approaches with the runtime efficiency of filter-based methods.