PaperlayerView exact revision on arXiv

Source-linked AI summary

A Unified Neural-Aided Alignment and Calibration Method for AUVs

Guy Damari, Zeev Yampolsky, Itzik Klein

arXiv:2608.22496v1cs.RO

TL;DR

AUV DVL initialization traditionally relies on model-based alignment and calibration that require demanding maneuvers, surface references, and simplified error models. This paper replaces both stages with ResAlignNet and DCNet operating on a single nearly constant-velocity trajectory. On real data across five error-term combinations, the pipeline reduces average VRMSE by 68.7% over the baseline using 25s of initialization data.

  • Problem

    Conventional DVL initialization requires complex maneuvers, surface-level satellite references, and simplified error models, making the process time-consuming and trajectory-dependent.

  • Method

    The pipeline uses ResAlignNet for alignment and DCNet for calibration, consuming velocity windows on a single nearly constant-velocity trajectory.

  • Results

    68.7%: the pipeline reduces average velocity root mean squared error over the model-based baseline using 25s of initialization data.

  • Takeaways & Limitations

    The unified neural-aided pipeline enables in-situ initialization without external positioning or dedicated maneuvers while retaining stable accuracy across five error-term combinations.

Abstract

from arXiv · show

Autonomous underwater vehicles (AUVs) rely on the fusion of inertial navigation systems (INS) and Doppler velocity logs (DVL) for accurate navigation. Before deployment, this fusion requires a DVL initialization pipeline consisting of two stages: alignment, which estimates the rotation between the INS and DVL frames, and calibration, which estimates the DVL error terms. Conventionally, both stages are solved with model-based algorithms that demand complex vehicle maneuvers, surface-level satellite reference measurements, and simplified error models, making initialization time-consuming, trajectory-dependent, and sensitive to sensor quality. In this work, we propose a fully neural- aided DVL initialization pipeline that replaces both stages with two complementary neural networks: ResAlignNet for alignment and DCNet for calibration. The unified pipeline operates in situ on a single nearly constant-velocity trajectory and uses the same inputs as the model-based baseline. Using real-world data recorded across five distinct sensor error-term combinations, the proposed pipeline reduces the velocity root mean squared error by an average of 68.7% over the model-based baseline, using only 25s of data for initialization.

I. INTRODUCTION

AUV DVL initialization traditionally combines alignment and calibration through model-based methods that require demanding operating conditions. The proposed neural pipeline learns both stages from data and achieves strong accuracy with a short, simple trajectory.

  • DVL initialization aligns INS and DVL frames, then estimates deterministic DVL scale-factor and bias errors.
  • Conventional alignment requires specific excitation maneuvers, long convergence times, or external positioning infrastructure.
  • The proposed approach learns both alignment and calibration directly from data, without analytical vehicle models, prescribed maneuvers, or external infrastructure.
  • 68.7%: the proposed pipeline reduces average velocity root mean squared error over the model-based baseline using 25s of data.

A. INS/DVL Alignment

INS/DVL alignment estimates the rotation between the body and DVL frames by matching inertial and DVL-derived velocities. The model-based formulation is accurate only with sufficient trajectory excitation and data, and degrades with lower-grade sensors.

  • Alignment estimates the rotation Td_b ∈ SO(3) between the body frame and the DVL frame.
  • The standard in-situ method integrates inertial acceleration into body-frame velocity and matches it with DVL velocity over N time instances.
  • SVD solves the velocity-residual minimization, but the baseline needs trajectories that sufficiently excite all axes and enough collected data.
  • Lower-grade sensors markedly degrade the SVD alignment baseline.

B. DVL Calibration Error Model

DVL calibration estimates deterministic measurement errors after alignment. The paper uses independent per-axis scale and bias terms rather than the scalar scale-factor assumption of the conventional baseline.

  • Calibration estimates deterministic DVL error terms relating measured DVL velocity to a reference velocity.
  • The adopted comprehensive model contains a scale-factor vector kDVL and a bias vector bDVL, each in R3.
  • Unlike scalar calibration, the six-term model applies an independent scale and bias to each axis for per-axis correction.

III. PROPOSED UNIFIED APPROACH

The unified approach replaces SVD alignment and scalar calibration with ResAlignNet and DCNet while retaining the same inputs. ResAlignNet regresses frame alignment from synchronized INS/DVL velocity windows.

  • ResAlignNet and DCNet replace SVD-based alignment and scalar scale-factor calibration with complementary neural estimators.
  • ResAlignNet uses synchronized six-channel INS and DVL velocity windows as input and outputs three Euler angles defining the estimated transformation.
  • The network trains alignment-angle predictions with mean squared error against ground-truth angles.
  • Because it operates directly on velocity windows, ResAlignNet requires neither prescribed maneuvers nor external positioning and converges within seconds of data.

B. Calibration Stage: DCNet

DCNet calibrates aligned DVL measurements by estimating per-axis scale and bias error terms through a multi-head neural network.

  • B. Calibration Stage: DCNet: DCNet estimates DVL error terms from aligned DVL and reference velocities expressed in the body frame.Its input stacks the DVL and reference velocity axes.
  • B. Calibration Stage: DCNet: The network uses separate convolutional heads to process velocity differences and jointly process corresponding axes.The two heads are concatenated before fully connected layers output the error terms.
  • B. Calibration Stage: DCNet: The six-term error model estimates an independent scale factor and bias for each axis.This provides more flexible per-axis correction than a scalar scale-factor model.
  • B. Calibration Stage: DCNet: DCNet is trained with mean squared error computed on calibrated velocity rather than directly on the error terms.The loss compares reference and calibrated velocities.

C. End-to-End Pipeline

The end-to-end pipeline chains ResAlignNet and DCNet: alignment transforms DVL measurements into the body frame, then calibration estimates per-axis scale and bias.

  • C. End-to-End Pipeline: ResAlignNet first estimates alignment from a calibration trajectory and transforms DVL measurements into the body frame.The resulting aligned DVL and reference velocities are passed to DCNet.
  • C. End-to-End Pipeline: DCNet then estimates the per-axis scale and bias, completing DVL initialization for subsequent trajectory correction.The resulting alignment and error terms are applied to evaluation trajectories.
  • C. End-to-End Pipeline: Both stages consume only velocity windows and operate on a single nearly constant-velocity trajectory.This removes the maneuver and surface-reference requirements of the model-based pipeline.

IV. EXPERIMENTAL RESULTS

The experiments use real AUV sea-trial data with separate calibration and evaluation trajectories to compare initialization performance across multiple error-term combinations.

  • IV. EXPERIMENTAL RESULTS: The evaluation uses six A-KIT trajectories totaling 33 minutes, with one for training, one for calibration, and four for evaluation.All trajectories contain synchronized DVL and inertial measurements plus a full navigation solution.
  • IV. EXPERIMENTAL RESULTS: The validation procedure estimates alignment and DVL error terms on one calibration trajectory before applying them to four independent evaluation trajectories.Evaluation trajectories are transformed to the body frame and calibrated.
  • IV. EXPERIMENTAL RESULTS: The procedure is repeated across five error-term combinations spanning alignment errors, per-axis scale factors, biases, and noise levels.Each combination is evaluated for both the model-based baseline and the proposed pipeline using a 25s initialization segment.

C. Accuracy Comparison

Across five error-term combinations, the proposed pipeline consistently outperforms the model-based baseline in average VRMSE using a 25s initialization segment.

  • C. Accuracy Comparison: 68.7% average improvement reduces VRMSE from 0.5768m/s for the baseline to 0.1803m/s for the proposed pipeline.The comparison averages results over four evaluation trajectories for each error-term combination.
  • C. Accuracy Comparison: The proposed pipeline outperforms the baseline across all five error-term combinations.Per-combination improvement ranges from 16.4% to 83.5%.
  • C. Accuracy Comparison: Under aggressive combinations such as ETC3, baseline VRMSE approaches 1.0m/s while the proposed pipeline remains below 0.21m/s.This comparison highlights differing performance under stronger sensor error conditions.

D. Ablation: Contribution of the Calibration Stage

Alignment provides most of the pipeline’s VRMSE reduction, while DCNet adds a smaller calibration refinement after accurate alignment.

  • Alignment alone reduces VRMSE far below the model-based baseline, identifying misalignment as the dominant error source.
  • 2.2%: DCNet lowers average VRMSE from 0.1843m/s to 0.1803m/s after ResAlignNet alignment.
  • DCNet improves four of five error-term combinations by up to 7.4%, but marginally increases error for ETC3.
  • Accurate alignment leaves residual DVL scale and bias errors small, limiting calibration to a modest refinement.

V. CONCLUSIONS

The paper presents a unified neural-aided DVL initialization pipeline that replaces conventional model-based alignment and calibration. On real AUV data, it improves average VRMSE substantially with only 25s of initialization data.

  • The unified pipeline chains ResAlignNet for alignment with DCNet for calibration, replacing the conventional SVD-and-scalar-scale procedure.
  • The evaluation uses real Snapir AUV data across five sensor error-term combinations.
  • 68.7%: the proposed pipeline reduces average VRMSE over the baseline using only 25s of initialization data.
  • The pipeline operates in situ without external positioning and uses a single nearly constant-velocity trajectory instead of dedicated maneuvers.
Loading 2608.22496v1…