Confidence-Gated Finite-Horizon Residual SAC–LSTM Control for Visual Target Tracking by a Rudderless Six-Thruster AUV

Autonomous visual tracking of a moving target by a rudderless underwater vehicle is complicated by nonlinear dynamics, intermittent visual observations, environmental currents, and constrained distributed thrust. This paper combines a recurrent SAC–LSTM policy with a finite-horizon model-based residual correction. The policy generates a baseline three-component virtual action, while the model branch predicts the terminal image error and computes a residual correction before constrained six-thruster allocation. A confidence gate attenuates or disables this correction when the target observation, depth estimate, or local control sensitivity is unreliable. The method was evaluated on a nominal five-degree-of-freedom model of the MMT-300 AUV. In a twenty-seed severe-condition experiment, the correction reduced the mean all-frame tracking RMSE from 1.289 to 0.897 and increased mean target visibility from 39.22% to 57.94%. This improvement was accompanied by an increase in the fraction of control cycles with at least one active thrust limit, from 53.1% to 55.3%. Across the complete six-method factorial grid, the recurrent policy and the full method had the same mean all-frame RMSE of 1.363, indicating that the benefit is concentrated in difficult operating regimes rather than being universal. A separate ablation also showed a consistent advantage of water-relative over ground-relative velocity in the model-based prediction under nonzero current.