DMIAN: Deep Multi-IMU Aided Navigation

Estimated results from the proposed method. Left: velocity over time. Right: position over time.

4.3 ms
Inference time (CPU)

0.11
Vel. RMSE (m/s)

0.88
R² Score

0.804
Pos. RMSE (m)

54.6%
Improvement vs INS-DVL

27.5%
Improvement vs NN baseline

Abstract

Learned inertial odometry has advanced rapidly across domains, especially in GNSS-denied environments. This paper introduces a learning-based approach that combines multiple IMUs with a DVL to improve velocity estimation for marine vehicles. The proposed method employs a multi-head attention Long Short-Term Memory network to fuse temporally and spatially distributed inertial signals with aiding velocity measurements. The model outputs both velocity estimates and their corresponding covariances, which are integrated as measurement updates within an EKF. This hybrid design allows learned features to complement traditional state estimation while maintaining filter consistency. The system is implemented and validated on the H2OmniX platform through a diverse set of trajectories. The method takes less than 5 ms for inference both on the GPU and the CPU, demonstrating less than 0.11 m/s RMSE and more than 0.88 R² in unseen trajectories through all ablation studies. The multi-IMU and DVL fusion provides the most accurate results, whereas the models with other IMU configurations continue to deliver reliable estimations when additional data are unavailable.

Deep Learning Inertial Navigation Multi-IMU LSTM ASV DVL EKF

Contributions

A data-driven method that accurately predicts vehicle velocities and their covariances through the fusion of a multi-IMU setup and additional aiding sensors, enabling direct integration within a standard EKF.
A multi-head attention LSTM architecture where each sensor branch independently extracts temporal motion features before fusion into velocity and uncertainty output heads.
Flexible fusion that handles any subset of available sensors at inference time, providing reliable estimations even when DVL or individual IMUs are unavailable.
Extensive experimental validation on the H2OmniX ASV, including ablation studies across all IMU configurations and comparison against other baselines.

System Design

H2OmniX ASV with the multi-IMU enclosure and DVL mounted underneath.

Platform. The H2OmniX is an omnidirectional ASV developed at LABUST. It measures 1 m in length, weighs approximately 15 kg, and runs navigation software on ROS2 Humble.

Sensor setup. Three Xsens MTi-630R IMUs are mounted orthogonally in a perpendicular triad inside an IP-rated enclosure at 100 Hz. A NavQuest 600 Micro DVL is mounted beneath the platform providing 3-axis velocity at 5 Hz.

EKF. A constant-acceleration motion model with state vector x = [p, Θ, v, ω, a]. All configurations share the same prediction model and process noise Q.

Method

Raw IMU and DVL data are transformed to a common body frame, then passed to the multi-head LSTM. Predicted velocities and diagonal covariances serve as EKF measurement updates.

System workflow from raw sensor data to EKF state variables.

Architecture. Each of the three IMU branches and the DVL branch is an independent 2-layer LSTM with 64 hidden units. Branch outputs are concatenated and passed through fully connected fusion layers into a velocity head and an uncertainty head.

Loss. Two-stage training: 10 epochs of MSE to stabilize velocity regression, then Gaussian Negative Log-Likelihood (GNLL) to jointly learn per-axis uncertainty. The diagonal covariance structure ensures a positive-definite measurement noise matrix for EKF integration.

Results

Comparison of DMIAN against BeamsNet and classical INS-DVL across five evaluation trajectories.

2D trajectory comparison across all five evaluation trajectories. Rows: INS-DVL (top), BeamsNet (middle), DMIAN (bottom).

Table 1 — Average Position Performance

Metric	BeamsNet	INS-DVL	DMIAN
MAE (m)	0.82632	1.20609	0.73442
MSE (m²)	0.87767	2.52672	0.74459
RMSE (m)	0.91023	1.50836	0.80424

Table 2 — Average Velocity Performance

Metric	BeamsNet	INS-DVL	DMIAN
MAE (m/s)	0.11640	0.18817	0.09710
MSE (m²/s²)	0.02389	0.06481	0.01252
RMSE (m/s)	0.15368	0.24539	0.11137
R²	0.75694	0.29777	0.87802

Position error over time — Trajectory 4.

Velocity estimates — Trajectory 4.

Position estimates — Trajectory 4.

Ablation Study

IMU count is varied from 1 to 3 while keeping architecture, training, and process noise Q identical. 1-IMU and 2-IMU results are averaged across all possible IMU combinations.

Velocity and Position Performance by IMU Configuration

Table 3 — Average Velocity (all with DVL)

Metric	1 IMU + DVL	2 IMU + DVL	3 IMU + DVL
MAE (m/s)	0.10510	0.10154	0.09587
MSE (m²/s²)	0.01481	0.01350	0.01222
RMSE (m/s)	0.12091	0.11572	0.10998
R²	0.85781	0.86895	0.88017

Velocity RMSE distribution per IMU configuration.

Position error distribution per IMU configuration.

Cumulative distribution of drift rate.

Cumulative distribution of position RMSE.

Citation

@article{batos2026dmian,
  title   = {DMIAN: Deep Learning-Based Multi-IMU Fusion for Enhanced Marine Aided Navigation},
  author  = {Bato\v{s}, Matko and Na\dj{}, \DJ{}ula},
  journal = {Submitted to Control Engineering Practice},
  year = {2026},
  volume = {},
  pages = {},
  issn = {},
  doi = {},
  url = {}
}

Citation details will be updated upon article publication.

Contact

Matko Batoš

matko.batos@fer.unizg.hr

Laboratory for Underwater Systems and Technologies (LABUST)
Faculty of Electrical Engineering and Computing, University of Zagreb
Unska ul. 3, 10000 Zagreb, Croatia