H-SPAR: Hydrodynamic-aware Simulation for Particle Transport and Autonomous Robots

Can a single simulation framework jointly model spatially and temporally varying currents, USV motion, and Lagrangian particle transport so that planning, execution, and sampling performance can be evaluated together?

Environmental sampling robots must navigate in currents that both push the vehicle and carry target particles, but existing marine simulators treat flow, robot autonomy, and sampling targets in separate or simplified ways, so mission designers cannot see how path, energy, and particle capture trade off before deployment.

Prior simulators such as DAVE/UUV Simulator, VRX, MOOS-IvP, MARUS, and UNav-Sim model some currents and vehicle disturbance, but none natively provide solver-derived fine-grained spatiotemporal velocity fields, Lagrangian particle transport, or probabilistic particle sampling. Current-aware planners VF-RRT* and SVF-RRT* optimize against static or time-averaged point-mass flow representations and ignore rigid-body hydrodynamic execution effects; coverage studies often assume uniform currents.

H-SPAR uses a precomputed fine-grained velocity field from shallow-water-equation solvers as a shared input. In ROS 2 it interpolates that field for both particle advection and hydrodynamic drag on a catamaran USV; Gazebo executes the vehicle dynamics with the VRX thrust controller, while RRT*, VF-RRT*, and SVF-RRT* generate waypoints and DWA handles low-level motion. A probabilistic sampling module uses a Gaussian capture probability centered on a 0.5 m/s design intake velocity to decide particle ingestion based on relative particle-robot speed.

In planning/execution benchmarks, RRT* produced the shortest planned path (348.2 m) but the highest planned upstream cost (1238.98); VF-RRT* and SVF-RRT* had costs of 667.71 and 379.49, respectively, and SVF-RRT* achieved 69.4% lower planned upstream cost than RRT*. Under time-varying execution, executed upstream costs were RRT* 1312.8, VF-RRT* 1054.2, SVF-RRT* 765.2, reducing SVF-RRT*'s advantage to 41.7%; SVF-RRT* also showed larger trajectory deviation (6.7 m vs RRT*'s 2.82 m). Under the time-averaged flow field used in planning, current-aware planners had lower mission times (SVF-RRT* 993.96 s vs RRT* 1230.97 s) with comparable deviation. In coverage tests with 100 particles released every 10 s, the best sweep orientation depended on modeling assumptions: under uniform deterministic flow the 90° sweep had the highest rate (0.46 particles/s), but under the full H-SPAR probabilistic configuration the 0° sweep had the highest rate (0.11), and rates varied by up to 22.2% across orientations.

The authors state that the results are simulation-only and have not been validated against physical particle transport or vehicle drift data; the flow field is depth-averaged shallow water, so vertical profiles, 3D turbulence, and near-body interactions are absent. H-SPAR uses one-way coupling, meaning the vehicle does not affect the flow or particles, and flow fields are precomputed rather than updated online. The path planners use time-averaged currents and static costmaps, and execution was tested on a single class of USV in Gazebo with DWA, so real-world transfer and other controllers are not demonstrated.