Abstract
Under degraded vision, docking requires a robust estimate of the relative dock state, but the capture window is short, and vision and sonar suffer from attenuation and ambiguity, respectively. In this simulation study, we develop a lightweight camera–forward-looking sonar (FLS) transformer and evaluate its interaction with terminal estimation and capture. At the strongest tested optical degradation, the transformer achieved 0.3079 m position RMSE versus 0.4251 m for capacity- and loss-matched late fusion. Extending terminal training support reduced near-field extrapolation error. A separate 880-trial BlueROV2 Heavy experiment without the transformer yielded an inconclusive primary target-geometry comparison. In a distinct 440-trial realistic-FLS simulation study, adding a Doppler velocity log (DVL) increased capture from 87/220 (39.55%) to 220/220 (100%). Cue ablation revealed dependence on artificial guidance beams, motivating adaptation to a markerless open-bore dock. A held-out evaluation after checkpoint selection of the adapted model obtained 0.06644 m position RMSE over 222 configurations. The integrated demonstrator docked 9/9 times; continuous transformer control, fixed terminal handover, and the integrated stack each docked 6/6 times in matched rich-world conditions. These bounded demonstrations establish simulated operability, not population reliability. Lower perception error helps, but successful capture also depends on sufficient terminal-state observability and motion propagation. Environment dependence and uncertainty calibration remain priorities for physical deployment.