# Attention-Augmented LSTM Feed-Forward Compensation for Lever-Arm-Induced Velocity Errors in Transfer Alignment

## TL;DR

This paper introduces a new method using an attention-based LSTM to reduce velocity errors in underwater robotic systems caused by changing lever arms during transfer alignment.

## Contribution

The novel approach combines an adaptive Kalman filter with an attention-based LSTM for real-time compensation of lever-arm-induced velocity errors.

## Key findings

- The proposed method reduces RMS misalignment angle error by 64% in underwater robot simulations.
- RMS installation error angle is reduced by 66% under typical maneuvers like acceleration and turning.
- The method improves robustness and practicality of transfer alignment with time-varying lever arms.

## Abstract

In a mother–child underwater bio-inspired robotic system, the equivalent lever arm between the master and slave inertial navigation systems (INSs) varies with launcher attitude changes and structural flexure. This time-varying lever arm introduces hard-to-model systematic velocity errors that degrade the accuracy and filter convergence of velocity difference-based transfer alignment. Traditional rigid body compensation relies on precise, constant lever-arm parameters and fails when booms, launch tubes, or flexible manipulators undergo appreciable deformation or reconfiguration. To address this, we augment a “velocity–attitude joint matching and innovation-based adaptive Kalman filter (AKF)” framework with an attention-based Long Short-Term Memory (LSTM) feed-forward module. Using only a short, real-time Inertial Measurement Unit (IMU) sequence from the slave INS, the module predicts and compensates the velocity bias induced by the lever arm. Numerical simulations of an underwater bio-inspired robot deployment scenario show that, under typical maneuvers (acceleration, turning, fin-flapping, and S-curve), the proposed method reduces the root-mean-square (RMS) misalignment angle error from about 14.5′ to 5.2′ and the RMS installation error angle from 8.8′ to 3.0′—average reductions of about 64% and 66%, respectively—substantially improving the robustness and practical applicability of transfer alignment under time-varying lever arms and flexible disturbances.