Inertial Navigation Meets Deep Learning: A Survey of Current Trends and Future Directions

TL;DR

This survey addresses how deep learning enhances inertial navigation and sensor fusion across land, aerial, and maritime domains. It distinguishes approaches that directly regress navigation states from those that refine sensors or filter parameters, and it catalogs architectures (MLP, CNN, RNN, transformers) and end-to-end versus residual strategies. Key contributions include a structured review by platform, a synthesis of calibration/denoising methods, and insights into trends such as increasing emphasis on learning filter parameters and adopting advanced architectures. The study underscores the potential of DL to improve GNSS-denied navigation and robust multi-sensor fusion, with implications for real-time, cross-domain autonomous systems.

Abstract

Inertial sensing is used in many applications and platforms, ranging from day-to-day devices such as smartphones to very complex ones such as autonomous vehicles. In recent years, the development of machine learning and deep learning techniques has increased significantly in the field of inertial sensing and sensor fusion. This is due to the development of efficient computing hardware and the accessibility of publicly available sensor data. These data-driven approaches mainly aim to empower model-based inertial sensing algorithms. To encourage further research in integrating deep learning with inertial navigation and fusion and to leverage their capabilities, this paper provides an in-depth review of deep learning methods for inertial sensing and sensor fusion. We discuss learning methods for calibration and denoising as well as approaches for improving pure inertial navigation and sensor fusion. The latter is done by learning some of the fusion filter parameters. The reviewed approaches are classified by the environment in which the vehicles operate: land, air, and sea. In addition, we analyze trends and future directions in deep learning-based navigation and provide statistical data on commonly used approaches.

Figures (9)

• Figure 1: (a) Strapdown inertial navigation algorithm. For a given initial condition, the gyroscope's angular velocity and the accelerometer's specific force measurements are integrated over time to calculate the navigation solution in the desired reference frame (local, geographic, and so on). (b) The process by which the navigation solution can be corrected using a nonlinear filter and an aiding sensor. By looking at the output, the black diverging curve shows how the navigation solution accumulates error and the red curve illustrates how the aiding sensor corrects this error, resulting in a chainsaw-like signal.
• Figure 2: The figure illustrates an LSTM architecture based on the DeepVIP architecture described in zhou2022deepvip. The DeepVIP architecture involves passing inertial readings and additional sensor data through LSTM cells to capture time dependencies. Subsequently, the data traverses dropout layers to prevent overfitting before being processed by FC layers to extract velocity and heading residual outputs.
• Figure 3: The figure illustrates the architecture presented in al2019deep, which is based on a simple MLP network. This network utilizes the estimated attitude states from the Kalman filter and enhances them through training with data containing accurate reference attitude information.
• Figure 4: The figure depicts the architecture based on A-KIT introduced in cohen2024kit, where an adaptive Kalman-informed transformer is employed. In this approach, inertial data along with observable states, such as position from GNSS, are fed through the block diagram presented. The output obtained from this process provides the scale factors required for dynamically estimating the process noise covariance matrix of the extended Kalman filter.
• Figure 5: The figure is based on the architecture introduced in chen2018improving, where a CNN architecture is proposed to address IMU errors by analyzing a window of inertial measurements, enabling the detection and removal of noisy features. The block diagram depicts noisy data passing through convolutional layers, with subsequent smoothing or filtering of the input.