Uncertainty-Aware GRU-Enhanced INS for Robust Navigation in AI-based Autonomous Vehicles

초록

In autonomous vehicle navigation, Inertial Navigation Systems (INS) provide reliable motion information but suffer from drift and accumulated errors when operating without Global Navigation Satellite System (GNSS) support. This paper proposes a Gated Recurrent Unit (GRU)-based deep learning framework to enhance yaw-rate estimation, which is essential for lateral localization of ground vehicles. The model predicts yaw rate from raw IMU measurements using GRU’s temporal modeling capability and evaluates prediction reliability through variance-based uncertainty analysis. A multi-head prediction structure is adopted to generate parallel outputs and dynamically select the most reliable estimate. During training, high-precision labels from GNSS/INS fusion are used, whereas inference relies solely on INS data and GRU outputs, enabling performance augmentation for low-cost INS sensors. Unlike conventional deep learning-Kalman filter hybrid approaches, the proposed method employs a direct RNN-based prediction structure, achieving simplicity and computational efficiency suitable for real-time embedded systems. The model is trained and evaluated using asynchronous accelerometer and gyroscope data from the KITTI OXTS dataset, with performance compared against ground-truth yaw angles from high-accuracy GNSS/INS fusion. Experimental results show that the proposed framework significantly reduces yaw-angle estimation error over traditional single-gyro integration, particularly in GNSS-denied dead-reckoning scenarios. These findings confirm the effectiveness of GRU-based temporal modeling in mitigating INS drift and enhancing navigation performance in autonomous vehicles. Copyright© (2026) by Institute of Navigation. All rights reserved.