AUV-Fusion: Underwater Sensor Fusion

• AUV-Fusion is a multi-modal framework that integrates heterogeneous sensors, such as INS and DVL, using nonlinear EKF techniques for improved underwater navigation.
• It combines classical filtering methods with AI-adaptive calibration and graph-based smoothing to enhance odometry and mapping accuracy.
• The framework addresses challenges like sensor outages, low visibility, and synchronization constraints, enabling resilient performance in complex underwater environments.

Searching arXiv for papers using the term “AUV-Fusion” and closely related underwater sensor-fusion frameworks. arxiv_search(query="AUV-Fusion autonomous underwater vehicle sensor fusion", max_results=10, sort_by="relevance") AUV-Fusion is a label used in recent arXiv literature for several, partly overlapping lines of work on autonomous underwater vehicle sensor fusion. In its narrow sense, it denotes an INS/DVL navigation framework that augments a nonlinear error-state EKF with DVL-derived acceleration updates to improve convergence and bias estimation (Levy et al., 2023). In a broader sense, it names a research program spanning classical model-based filtering, AI-adaptive navigation, tightly coupled visual–acoustic–inertial odometry, and opti-acoustic volumetric mapping for AUVs operating without GNSS and under visibility, synchronization, and sensor-availability constraints (Damari et al., 6 May 2026, Collado-Gonzalez et al., 15 Mar 2026). The same name is also used in an unrelated cross-modal adversarial attack framework for visual-aware recommender systems, which is a distinct topic and not part of underwater robotics (Ling et al., 30 Jul 2025).

1. Terminological scope and research setting

Within underwater robotics, AUV-Fusion refers to the integration of heterogeneous sensing modalities—most often INS, DVL, cameras, sonar, pressure sensors, magnetic compasses, LBL, GNSS surface fixes, and IMU-derived preintegration—into a single estimation or mapping pipeline. The stated motivation is consistent across the literature: electromagnetic signals are unavailable underwater, DVL measurements may become incomplete or unavailable, visual sensing is highly sensitive to turbidity and low contrast, and asynchronous sensors complicate tightly coupled estimation (Damari et al., 6 May 2026, Cohen et al., 2024, Collado-Gonzalez et al., 15 Mar 2026).

The term has both framework-specific and umbrella usages. In “INS/DVL Fusion with DVL Based Acceleration Measurements,” the integrated navigation method itself is described as “AUV-Fusion” and is centered on software-only modification of an INS/DVL EKF (Levy et al., 2023). In “AI-Aided Advancements in Autonomous Underwater Vehicle Navigation,” “AUV-Fusion” is used as a compendium-level designation for a taxonomy of classical filters, learning-based calibration, adaptive filtering, visual–inertial fusion, and emerging reinforcement-learning-driven sensor weighting (Damari et al., 6 May 2026). This suggests that the term functions less as a single standardized algorithmic object than as a family resemblance across underwater fusion architectures.

A concise cross-section of the literature is given below.

Paper	Modalities	Principal focus
(Levy et al., 2023)	INS + DVL	DVL-based acceleration update inside EKF
(Cohen et al., 2024)	INS + DVL	Missing-beam regression with HNC for LC/TC fusion
(Zhao et al., 2023)	Vision + DVL + IMU + Pressure	Tightly coupled MSCKF odometry
(Song et al., 2023)	SINS + LBL + DVL + MCP + PS + GNSS	Factor-graph integrated navigation
(Collado-Gonzalez et al., 15 Mar 2026)	Stereo sonar + monocular camera	Confidence-weighted GP volumetric mapping
(Westerdahl et al., 4 Dec 2025)	LiDAR + IMU + dual orthogonal FLS	Seabed-to-sky factor-graph mapping

2. INS/DVL-centered navigation formulations

A core AUV-Fusion lineage is the INS/DVL error-state filter. The canonical state comprises position, velocity, attitude or misalignment, and inertial biases, while DVL velocity enters as an aiding measurement. In the acceleration-augmented variant, the error-state is

$$\delta x = \bigl[\;\delta v^n\,,\,\phi^n\,,\,b_a\,,\,b_g\;\bigr]^T \in\mathbb R^{12},$$

with continuous-time linearized dynamics

$\dot{\delta x} = F\,\delta x + G\,w.$

The distinctive step is the introduction of a DVL-derived acceleration measurement obtained from a least-squares fit over recent DVL velocity samples. With

$V = (L^T L)^{-1} L^T\,V_h,$

the acceleration estimate is extracted from the second row, and the EKF is updated using

$$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$

The reported effect is not the creation of new observable modes—the paper explicitly states that the unobservable null-space is unchanged—but faster convergence of accelerometer and gyroscope residual errors and lower residual bias error (Levy et al., 2023).

This INS/DVL-centered thread has diversified in several directions. A hybrid adaptive velocity-aided navigation filter learns the momentary process-noise covariance $Q_k$ from handcrafted IMU features and inserts the learned covariance into an es-EKF, reporting SRMSE $0.980$ m/s and SMAE $0.866$ m/s for the hybrid HCF-Ensemble, compared with $1.082$ m/s and $0.961$ m/s for a constant nominal $Q$ baseline (Or et al., 2022). A DVL outage formulation, ST-BeamsNet, regresses body-frame velocity during complete DVL outage from inertial windows and previous DVL velocity history using a Set-Transformer, reporting RMSE $\dot{\delta x} = F\,\delta x + G\,w.$0 m/s, MAE $\dot{\delta x} = F\,\delta x + G\,w.$1 m/s, $\dot{\delta x} = F\,\delta x + G\,w.$2, VAF $\dot{\delta x} = F\,\delta x + G\,w.$3, and an $\dot{\delta x} = F\,\delta x + G\,w.$4 speed RMSE, which is approximately $\dot{\delta x} = F\,\delta x + G\,w.$5 better than the moving-average approach (Cohen et al., 2022). A complementary missing-beam framework, HNC, regresses two or three absent Janus beams and reinserts them into loosely or tightly coupled EKF updates; for two missing beams it reports VRMSE $\dot{\delta x} = F\,\delta x + G\,w.$6 m/s on both trajectories for HNLC and HNTC, and for three missing beams $\dot{\delta x} = F\,\delta x + G\,w.$7 m/s and $\dot{\delta x} = F\,\delta x + G\,w.$8 m/s, with average performance gains reported as $\dot{\delta x} = F\,\delta x + G\,w.$9 relative to baseline model-based approaches and $V = (L^T L)^{-1} L^T\,V_h,$0 relative to a model-based beam estimator (Cohen et al., 2024).

A separate but adjacent issue is INS/DVL alignment. AlignNet formalizes alignment as direct regression from synchronized $V = (L^T L)^{-1} L^T\,V_h,$1 windows to Euler-angle offsets $V = (L^T L)^{-1} L^T\,V_h,$2. On synthetic lawn-mower trajectories, it reports RMSE $V = (L^T L)^{-1} L^T\,V_h,$3 at $V = (L^T L)^{-1} L^T\,V_h,$4 s, $V = (L^T L)^{-1} L^T\,V_h,$5 at $V = (L^T L)^{-1} L^T\,V_h,$6 s, $V = (L^T L)^{-1} L^T\,V_h,$7 at $V = (L^T L)^{-1} L^T\,V_h,$8 s, and $V = (L^T L)^{-1} L^T\,V_h,$9 at $$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$0 s, with convergence to approximately $$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$1 RMSE in approximately $$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$2 s versus approximately $$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$3 s for the velocity-based SVD baseline (Damari et al., 27 Mar 2025). This suggests that, in the AUV-Fusion literature, calibration and estimator adaptation are treated as integral parts of fusion rather than merely preprocessing.

3. Tightly coupled odometry and smoothing back-ends

A second major branch of AUV-Fusion emphasizes tightly coupled multi-sensor odometry. In the under-ice visual–DVL–IMU–pressure framework, visual features, DVL velocity, IMU mechanization, and pressure depth are integrated within a Multi-State Constraint Kalman Filter. The state is

$$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$4

with adaptive keyframe cloning and a DVL-aided feature enhancement that uses sparse DVL point clouds to correct visual feature depth under short baselines. In a frozen Keweenaw Waterway dataset of about $$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$5 m, plain MSCKF plus vision, DVL, and pressure achieved RMSE$$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$6 approximately $$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$7 m, adding keyframe clones reduced it to approximately $$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$8 m, and the full system produced best ATE$$\delta z_a = \tilde a^b \;-\; R^b_d\,\hat a^d, \qquad H_a = \bigl[\,0_{3\times3}\;\;-\;R^b_n\,[g^n\times]\;\;I_3\;\;0_{3\times3}\bigr].$$9 approximately $Q_k$0 m, while monocular VIO alone failed under prolonged hovering and DVL-IMU-pressure dead reckoning drifted unbounded (Zhao et al., 2023).

Factor-graph smoothing provides a distinct fusion regime. FGO-ILNS represents each state as

$Q_k$1

and constructs a sliding-window graph containing IMU preintegration factors, floating-LBL slant-range difference factors, GNSS, DVL, MCP, PS, and a marginalization prior. Asynchronous sensors are handled by forward-backward IMU preintegration, and history outside the window is compressed by Schur complement. The optimization solves

$Q_k$2

using Gauss–Newton or Levenberg–Marquardt in Ceres Solver. In simulation, the best ENU RMS at a $Q_k$3 s sliding window was East $Q_k$4 m, North $Q_k$5 m, Up $Q_k$6 m; under aggressive maneuvers, EKF drift exceeded $Q_k$7 m while FGO stayed below $Q_k$8 m; and in an outage scenario the maximum drift remained below $Q_k$9 m for FGO versus above $0.980$0 m for a federated EKF (Song et al., 2023).

FAR-AVIO advances the tightly coupled formulation by embedding a Schur-complement landmark elimination inside an EKF. Its nominal state augments pose, velocity, inertial biases, and both camera and DVL extrinsics,

$0.980$1

and the minimal error state has $0.980$2. Visual updates are reduced to fixed-size equivalent observations by eliminating landmark states through

$0.980$3

which yields constant-time EKF updates. The framework adds AWARE, an online reliability mechanism that scales measurement covariances and can disable or re-enable a sensor stream based on quality queues and thresholds, and it performs online DVL–IMU extrinsic calibration by including $0.980$4 in the state. On eight tank sequences, FAR-AVIO reports overall RMSE approximately $0.980$5 m versus $0.980$6 m for AQUA-SLAM and more than $0.980$7 m for ORB-SLAM3 and VINS-Fusion; AWARE improves RMSE by $0.980$8–$0.980$9; online extrinsic calibration converges to below $0.866$0 rad and below $0.866$1 m; and runtime on Jetson Orin NX is $0.866$2 ms/frame versus $0.866$3 ms for VINS-Fusion (Wei et al., 23 Dec 2025).

4. Opti-acoustic fusion and volumetric mapping

Another AUV-Fusion trajectory moves from state estimation to geometry reconstruction. “Towards Versatile Opti-Acoustic Sensor Fusion and Volumetric Mapping” proposes a BlueROV2 platform equipped with a forward-looking monocular camera and two orthogonally mounted multibeam imaging sonars: a “horizontal” sonar with $0.866$4 H $0.866$5 $0.866$6 V and a “vertical” sonar with $0.866$7 H $0.866$8 $0.866$9 V. Overlapping sonar fan-planes resolve elevation ambiguity in the intersection region, while the camera provides elevation cues through ROI segmentation using YOLO11n-seg. Stereo-sonar points, sonar-to-image projection points, and image-expansion points are each assigned confidence weights $1.082$0, $1.082$1, and $1.082$2, with heteroscedastic GP noise

$1.082$3

Occupancy is then estimated by Gaussian Process Volumetric Mapping with a Matérn $1.082$4 kernel,

$1.082$5

and converted to occupancy probability via

$1.082$6

The reported voxel resolution is $1.082$7 cm, keyframes are triggered every $1.082$8 cm or $1.082$9, confidence-weighted GP mapping adds approximately $0.961$0 s/frame relative to standard GP, and OctoMap is more than $0.961$1 slower. In tank experiments, the proposed GPC SS RGB achieved approximately $0.961$2 cm MAE for a single disk and $0.961$3 cm for a double disk, together with the lowest RMSE and precision above $0.961$4; in a turbid marina, it qualitatively captured wooden pilings and small front-pipe features, while GP-only and OctoMap baselines either lacked coverage or overestimated occupancy (Collado-Gonzalez et al., 15 Mar 2026).

A related but surface-oriented extension fuses a downward-looking 3D LiDAR and IMU with a dual orthogonal forward-looking sonar pair in a modified LIO-SAM back-end. Stereo-derived 3D sonar points and leading-edge line scans are inserted into a single factor graph through motion-interpolated poses between LiDAR keyframes. The system reports approximately $0.961$5 Hz map updates and approximately $0.961$6 Hz odometry, above-water Euclidean errors of $0.961$7–$0.961$8 m after rigid alignment to UTM32, and underwater wall-normal cosine similarity of approximately $0.961$9–$Q$0, while noting a small systematic underestimation and lateral offset due to residual extrinsic errors and elevation uncertainty outside the stereo overlap (Westerdahl et al., 4 Dec 2025). Although this platform is an autonomous surface vehicle rather than an AUV, it extends the same orthogonal-sonar and factor-graph principles into unified maritime mapping.

5. Mathematical patterns across the literature

Across these works, AUV-Fusion is less a single estimator than a recurring set of mathematical design patterns. The most basic layer is model-based Bayesian filtering. The compendium literature distinguishes EKF, UKF, and PF as the canonical classical architectures. For EKF, the prediction and measurement update are written as

$Q$1

$Q$2

and adaptive weighting may update $Q$3 by covariance matching (Damari et al., 6 May 2026). This formulation underlies the INS/DVL acceleration update, missing-beam HNC, and several visual–DVL–inertial systems.

A second pattern is graph-based smoothing with preintegration and marginalization. FGO-ILNS uses forward-backward preintegration to project asynchronous sensors onto adjacent IMU states and applies a Schur-complement prior when sliding the window (Song et al., 2023). FAR-AVIO instead moves Schur elimination into the EKF update itself, achieving fixed state size and constant-time updates (Wei et al., 23 Dec 2025). This suggests that the boundary between “filter” and “optimizer” has become porous in modern underwater fusion.

A third pattern is learned adaptation or learned measurement completion. Examples include bagged-tree estimation of process-noise covariance from IMU statistics (Or et al., 2022), Set-Transformer regression of velocity during complete DVL outage (Cohen et al., 2022), HNC beam completion for limited DVL measurements (Cohen et al., 2024), and 1D-CNN alignment regression for $Q$4 estimation (Damari et al., 27 Mar 2025). The AI-aided survey additionally places these alongside ResAlignNet, A-KIT, ProcessNet, Gaussian-process covariance adaptation, PiDR, and prospective actor–critic tuning of filter covariances or sensor weights (Damari et al., 6 May 2026). A plausible implication is that AUV-Fusion research increasingly treats uncertainty tuning, alignment, and sensor availability as learning problems embedded around classical estimators rather than replacements for them.

6. Empirical behavior, limitations, and boundaries of the term

The reported performance gains in AUV-Fusion papers are substantial but method-specific. The acceleration-augmented INS/DVL EKF reports Z-axis accelerometer bias error reductions of $Q$5, X and Y gyro bias error reductions of $Q$6, leveling-angle reductions of $Q$7, average improvement of approximately $Q$8 in a straight run, and average improvement of approximately $Q$9 in a figure-eight experiment, with convergence-time reductions of $\dot{\delta x} = F\,\delta x + G\,w.$00 and $\dot{\delta x} = F\,\delta x + G\,w.$01, respectively (Levy et al., 2023). HNC reports seamless operation under two- or three-beam loss (Cohen et al., 2024). FGO-ILNS reports greater than $\dot{\delta x} = F\,\delta x + G\,w.$02 horizontal-accuracy improvement under complex dynamics relative to EKF and maintains less than $\dot{\delta x} = F\,\delta x + G\,w.$03 m maximum drift in outage scenarios (Song et al., 2023). Opti-acoustic GP mapping reports lowest MAE and RMSE among the compared baselines while maintaining substantial coverage in both clear and turbid water (Collado-Gonzalez et al., 15 Mar 2026).

At the same time, the limitations are explicit. The DVL acceleration update “does not introduce new observable modes” and “does not rectify classical unobservable biases (e.g. heading drift)” (Levy et al., 2023). AlignNet is evaluated on high-fidelity simulation and may face degradation from multipath acoustic effects and irregular seafloor in real deployment (Damari et al., 27 Mar 2025). The AI-aided overview emphasizes time synchronization, visibility and feature degradation in turbid water, computational constraints on onboard CPUs, generalization across vehicles and sensor grades, and the need to balance real-time accuracy with energy and latency (Damari et al., 6 May 2026). The dual-sonar/LiDAR mapping paper attributes systematic underwater offsets to residual extrinsic errors and elevation uncertainty outside stereo overlap (Westerdahl et al., 4 Dec 2025). These are not peripheral issues; they are recurrent structural constraints on underwater fusion.

Finally, the term’s boundaries matter. “AUV-Fusion: Cross-Modal Adversarial Fusion of User Interactions and Visual Perturbations Against VARS” is a recommender-systems attack framework that models user preference embeddings and injects perturbations into a diffusion-VAE latent space; despite the shared name, it does not concern autonomous underwater vehicles, inertial navigation, or marine sensing (Ling et al., 30 Jul 2025). For the underwater literature, AUV-Fusion therefore denotes a technical tradition centered on multi-sensor integration for navigation, odometry, and mapping, rather than a single universally defined algorithm.