Multi-sensor fusion localization framework using UKF-LSTM with Multiple Model Adaptive Estimation

doi:10.1016/j.robot.2026.105491

Multi-sensor fusion localization framework using UKF-LSTM with Multiple Model Adaptive Estimation

2026-04-28

0

PRE

AI

DOI：10.1016/j.robot.2026.105491

Original

Original request for help

Save

Abstract

En 中文

Accurate localization for mobile robots remains challenging under time-varying sensor uncertainty, where individual sensing modalities exhibit complementary but imperfect characteristics. Conventional filtering or learning-based approaches often struggle to maintain robustness when measurement quality varies over time. To address this issue, this paper proposes a hybrid multi-sensor fusion framework that combines learning-based error compensation with model-based state estimation. The method integrates LSTM-based correction of position observations with a dual-layer Multiple Model Adaptive Estimation (MMAE) structure and a UKF-based refinement mechanism, enabling adaptive handling of varying uncertainty conditions. Simulation results in a Gazebo environment show that the proposed approach consistently improves localization accuracy compared with standalone UKF and LSTM methods across multiple noise levels, achieving RMSE reductions ranging from 8.8% to 49.6%. The performance gain becomes more pronounced under moderate and severe noise conditions, where single-model approaches are less effective.

Keywords:

Localization

Multi-sensor fusion

LSTM

UKF

Adaptive estimation

Multi-sensor fusion localization framework using UKF-LSTM with Multiple Model Adaptive Estimation

Abstract

Journal

Organization

Cited Papers