Geometry-Aware Surrogate for Real-Time Hydrodynamics Estimation of Autonomous Ground Vehicles in Amphibious Environments
For autonomous ground vehicles operating in amphibious environments, this provides a computationally efficient way to embed physically accurate hydrodynamics into real-time simulation and planning loops.
This work presents a per-surface neural network surrogate that predicts geometry-resolved hydrodynamic forces for autonomous ground vehicles in real time (under 0.9 ms per sample), trained on CFD data. The model achieves a longitudinal-force sMAPE of 13% and vertical-force sMAPE of 3-12%, and reproduces physical relationships like quadratic drag scaling (R² ≥ 0.97) without explicit encoding.
Autonomous ground vehicles operating in shallow water or flood-prone terrains require dynamic models that account for hydrodynamic forces. However, the simulation and planning tools currently available either lack the physical fidelity or are too computationally expensive to run in real time. This work presents a per-surface neural network surrogate that bridges this gap by predicting geometry-resolved hydrodynamic forces at real-time rates, trained entirely on high-fidelity CFD data from two geometrically distinct vehicles. A vehicle specific Signed Distance Field (SDF) provides per-surface submergence inputs, allowing the model to resolve how loading varies with vehicle geometry, depth, and flow direction. On held-out CFD data, the surrogate achieves a longitudinal-force symmetric MAPE (sMAPE) of 13\% and a vertical-force sMAPE of 3-12\%, with inference running under 0.9\,ms per sample. To evaluate the model under real-world conditions, water wading trials of a full-scale vehicle at different submersion depths are used. Motion capture derived kinematics serve as the surrogate inputs, and the resulting predictions are tested to reproduce known physical relationships between force, speed, and depth. The predicted drag follows quadratic speed scaling ($R^2 \geq 0.97$) and the buoyancy intercepts scale linearly with depth ($R^2 = 0.973$). Neither relationship is encoded in the model training loss, both emerge from the per-surface architecture summing individually predicted surface forces. The resulting framework provides a pathway for embedding physically grounded hydrodynamics into the simulation and planning loops that autonomous ground vehicles depend on in amphibious environments.