Reliable underwater visual inspection in confined corridors and near-seabed environments demands attitude stabilization under boundary-induced disturbances, because small rotational jitter degrades image acquisition, feature tracking, image registration, and defect localization. This paper presents a disturbance-aware hierarchical control allocation framework for a four-fin bio-inspired AUV. The outer loop combines sliding-mode attitude control with a first-order disturbance observer; the inner loop maps desired body torques to bounded fin commands through regularized allocation. Unlike conventional designs that treat disturbance rejection and actuator allocation separately, the proposed framework explicitly retains the allocation residual in both the lumped disturbance model and a uniform ultimate boundedness analysis. Numerical evaluation on a reduced-order rotational model shows that the proposed method achieves the lowest disturbance-case RMS attitude error among PID, LQR, SMC, and ADRC baselines ( \(0.0267^\circ \) vs. \(0.0440^\circ \) for ADRC and \(0.1077^\circ \) for SMC), with strong robustness confirmed by a 200-run Monte Carlo study ( \(p < 10^{-66}\) ), a \(10\times \) IMU-noise sweep ( \(24\%\) degradation versus \(659\%\) for ADRC), and multi-scenario confined-environment tests. Sinusoidal tracking under continuous boundary disturbance achieves \(0.017^\circ \) RMS, supporting smooth inspection trajectories. In a combined tracking-plus-disturbance scenario where the controller must simultaneously settle a step command and reject the boundary disturbance, SMC and ADRC achieve slightly lower disturbance-phase RMS than the proposed method due to a transient observer–adaptation phase, which is acknowledged as a known limitation of the proposed architecture in this specific scenario. A motion-blur proxy analysis further demonstrates that the proposed attitude-error reduction translates into a \(39\%\) lower estimated blur index compared with ADRC and a \(75\%\) reduction compared with SMC, providing a direct quantitative link between control performance and visual-inspection quality. Simulation code and numerical data are openly available at https://github.com/NingCao1/AUVHierarchicalControl. Future work will validate the framework with real underwater imagery and task-level visual metrics including feature-match retention, image registration error, and defect-detection accuracy.
