Technical Summary: 3D Underwater Path Planning via Generative Flow Field Surrogates

Problem Statement

Autonomous Underwater Vehicle (AUV) Launch and Recovery (LAR) into the hull of an advancing host platform (e.g., an Extra-Large Uncrewed Underwater Vehicle, XLUUV) requires navigating a complex, three-dimensional propeller wake. This environment is characterized by localized velocity fluctuations, shear layers, and tip vortices that cannot be accurately modeled by uniform current fields. While high-fidelity Reynolds-Averaged Navier-Stokes (RANS) Computational Fluid Dynamics (CFD) simulations can resolve these structures with sufficient accuracy for path planning, their computational cost (hours per simulation) renders them impractical for onboard, real-time decision-making. Consequently, existing planners rely on simplified models (uniform currents or 2D approximations) that fail to capture the hydrodynamic loads critical to energy efficiency and control stability during close-proximity operations.

The central research gap addressed is twofold:

1. There is no systematic quantification of the performance loss incurred by ignoring 3D wake structures in AUV LAR planning.
2. It is unknown whether computationally tractable surrogate models can recover the benefits of full hydrodynamic knowledge within a planning loop.

Methodology

1. Hydrodynamic Domain and Dataset

The study utilizes a pseudo-static environment where the host XLUUV's speed and heading are held constant for each trajectory. The flow field is generated using RANS CFD simulations (ANSYS Fluent) with a realizable $k-\epsilon$ turbulence model and an actuator disk propeller model.

• Domain: A $155 \times 155 \times 155$ m volume centered on the XLUUV.
• Dataset: 550 distinct flow conditions spanning speeds of 0.10–5.00 m/s and heading angles of 0°–50°. These were interpolated onto a $128^3$ voxel grid (approx. 1.21 m resolution) to serve as ground truth and training data.

2. Path Planning Framework

An energy-weighted A* algorithm is employed on the $128^3$ voxel grid.

• Cost Function: The traversal cost between nodes is derived from the hydrodynamic drag force ($F_D \propto v^2$), where $v$ is the velocity magnitude at the destination node. This formulation incentivizes routing through low-velocity wake corridors (e.g., leeward margins) even if the geometric path is longer.
• Heuristic: An admissible heuristic based on the global mean energy cost per unit distance ensures optimality guarantees.

3. Generative Flow Field Surrogates (cGANs)

To replace the computationally expensive RANS simulations, the authors integrate two Conditional Generative Adversarial Network (cGAN) architectures into a hierarchical pipeline. The pipeline accepts scalar operating conditions (speed $v$ and heading $\theta$) as input and outputs a full $128^3$ volumetric flow field.

• Stage 1 (2D Synthesis): A pre-trained 2D cGAN (Model 4) synthesizes horizontal and vertical 2D planar velocity slices from the scalar metadata.
• Stage 2 (3D Generation): Two distinct 3D generators lift these slices to a full 3D volume:
  • Regularised PatchGAN GradNorm: A lighter architecture using 2D slice encoders, cross-attention, and gradient-normalized training. It is optimized for inference speed and memory efficiency.
  • 2D3DGAN with Self-Attention (2D3DGAN SA): A higher-fidelity architecture adapted from Yousif et al., featuring a global 3D self-attention layer to model long-range hydrodynamic relationships. It offers superior reconstruction metrics but requires significantly more GPU memory.

4. Experimental Design

The study benchmarks four planner variants across 19,800 independently generated trajectories (36 start nodes $\times$ 550 conditions):

1. Current-Only: Uses a uniform mean velocity field (naive baseline).
2. RANS CFD: Uses ground-truth interpolated CFD data (upper bound of performance).
3. PatchGAN: Uses the flow field predicted by the PatchGAN pipeline.
4. 2D3DGAN SA: Uses the flow field predicted by the 2D3DGAN SA pipeline.

Performance is evaluated using five metrics: Energy Expenditure, Path Length, High-Velocity Cell Encounters, Turbulent Cell Transitions, and Compute Time.

Key Contributions

1. Systematic Multi-Fidelity Benchmark: The first quantitative assessment of 3D path planning performance for AUV LAR, comparing uniform current, ground-truth CFD, and two cGAN surrogates across 19,800 trajectories. This establishes the "cost" of environmental ignorance and the recovery potential of surrogates.
2. Hierarchical cGAN Pipeline: Validation of a real-time, two-stage pipeline (2D slice synthesis $\to$ 3D volume generation) that generates full flow fields from scalar metadata in under 150 $\mu$s, making onboard deployment feasible.
3. Holistic Evaluation Framework: Introduction of a five-metric evaluation system that reveals a non-trivial trade-off: optimizing for energy efficiency (avoiding the high-velocity wake core) increases exposure to turbulent shear layers, a finding with direct implications for AUV control system design.

Results

Value of Hydrodynamic Knowledge

• Energy Savings: Full CFD knowledge reduces energy expenditure by 5.7–12.5% compared to uniform current planning. Savings scale with heading angle (asymmetry) rather than speed.
• Safety: CFD planning reduces encounters with high-velocity wake cores by up to 77.8%, significantly mitigating control saturation risks.
• Trade-offs: These benefits come at the cost of 1.9–8.5% longer path lengths and a 14–59% increase in turbulent cell traversals (exposure to shear layers).

Surrogate Performance

• Recovery of Benefits: Both cGAN surrogates recover approximately 45–60% of the CFD energy and high-velocity avoidance benefits at demanding operating conditions.
• Architectural Divergence:
  • 2D3DGAN SA: Demonstrates superior global reconstruction fidelity (lower MSE, higher SSIM) and maintains positive safety benefits (high-velocity avoidance) even at low heading angles. However, it requires ~25.7 GB of GPU memory, limiting it to pre-mission or surface-assisted planning.
  • PatchGAN: Offers a memory footprint of ~2.4 GB, suitable for embedded hardware (e.g., NVIDIA Jetson). While it performs well at high angles, it exhibits a critical failure mode at low heading angles: it under-predicts peak velocities in the compact propeller race, causing the planner to route into the high-velocity core, resulting in 121% more high-velocity encounters than the naive baseline.
• Reconstruction vs. Utility: The study finds that standard reconstruction metrics (MSE, SSIM) do not linearly predict downstream planning utility. A model with slightly worse global metrics (PatchGAN) can be operationally catastrophic if errors are concentrated in planning-critical regions (the wake core).

Computational Performance

• Inference Speed: The cGAN pipelines generate flow fields in 28–146 $\mu$s, which is four orders of magnitude faster than RANS simulation.
• Search Time: Interestingly, the GAN-informed planners exhibit faster A* search times than the CFD planner (44.9s vs 55.7s) because the smoother boundaries in GAN-predicted fields create a more navigable cost landscape, reducing the graph exploration required.

Significance and Claims

The paper claims to provide the first systematic quantification of the downstream path planning value of cGAN-predicted hydrodynamic fields in a 3D maritime robotics application.

• Operational Value: It demonstrates that 3D wake knowledge is not merely an incremental improvement but an operational necessity for energy efficiency and safety in close-proximity LAR, with benefits scaling with flow severity.
• Feasibility of Surrogates: It establishes that cGAN surrogates can recover a substantial fraction (45–60%) of the benefits of full CFD knowledge at inference speeds compatible with edge devices.
• Deployment Roadmap: The work delineates a clear path for deployment: the lightweight PatchGAN architecture is immediately deployable on current embedded hardware for real-time planning, while the higher-fidelity 2D3DGAN SA is suitable for pre-mission planning or future hardware generations.
• Methodological Insight: The study challenges the assumption that high global reconstruction accuracy guarantees good downstream task performance, arguing for training objectives that prioritize planning-relevant spatial features over global volume fidelity.

The authors conclude that while the current surrogates do not fully match the theoretical optimum of ground-truth CFD, they make real-time, wake-aware path planning tractable, quantifying the significant cost (in energy and safety margin) of operating without such knowledge.