DNN-Based Slip Ratio Estimator for Lugged-Wheel Robot Localization in Rough Deformable Terrains

Abstract

This paper presents a deep neural network (DNN)-based slip ratio estimator fused with an invariant extended Kalman filter (IEKF) for lugged-wheel robot localization using an inertial sensor and an encoder. Among various sensors used in wheeled mobile robot (WMR) localization, inertial sensors and encoders are most commonly used because these sensors are inexpensive and have low computational requirements. However, inertial sensors and encoders can cause large drifts in localization due to inherent sensor characteristics and wheel slippage, respectively. Most studies on wheel slippage have primarily focused on rubber tires, and using this slip ratio model for WMRs with lugged-wheels operating in outdoor environments can result in significant estimation errors in slip ratios. This paper develops a DNN-based slip ratio estimator and IEKF for WMR localization that is robust to wheel slippage even in rugged outdoor environments. The performance of the proposed localization is demonstrated through experiments using outdoor datasets where WMRs with lugged-wheels experience various slip conditions. Experiments are conducted in wet and dry conditions on a sloped grass field. Results show that the proposed localization method reduces accumulated localization errors by 53.5% compared to integration-based localization and by 13.5% compared to IEKF-based localization.