A safety-consistent hybrid global–local path planning framework for AUVs in bathymetry-constrained 3D marine environments

Autonomous underwater vehicles (AUVs) operating in marine environments must navigate rugged bathymetry, ocean-current disturbances, sudden obstacles, and limited perception. These conditions require long-range route guidance and short-range reactive avoidance, but inconsistencies can arise when planning stages use different clearance definitions, final supervision modifies local commands, or abrupt switching produces directional discontinuities. This paper proposes a planning-layer safety-consistent framework for bathymetry-constrained three-dimensional AUV navigation, in which a Dynamic Safety Domain (DSD) provides a shared clearance interpretation across planning and execution. The framework integrates the Hybrid Geometry-guided Artificial Lemming Algorithm (HG-ALA), the Geometry-Adaptive Softmax 3D Vector Field Histogram (GAS-3DVFH), and DSD-guided dynamic virtual sub-goal blending (DSD-GDVS) for global reference generation, local guidance, and continuous coordination, respectively. In the tested benchmarks, HG-ALA was the only population-based method to complete all 150 global-planning trials. In the dynamic-encounter benchmark, GAS-3DVFH retained greater physical clearance and required fewer supervisor contractions than 3D-DWA and RSO-VO. Factorized tests showed that DSD increased the mean retained safety margin in all matched comparisons, whereas continuous fusion reduced the mean directional-variation metrics. Their combined advantage was clearest in dynamic encounters.