A camera that hunts the wake instead of holding still: an agent finds hydrofoil wakes in a real water tunnel in about half a full scan's time
Synopsis
Particle image velocimetry (PIV) usually fixes the imaging position before the measurement, so if the structure of interest departs from expectation, no later algorithm can recover information that was never captured. This work treats camera placement as a closed-loop decision variable: a DreamerV3 agent trained in Isaac Sim drives a 6-DoF arm using only the velocity field and uncertainty estimated by LIMA to locate, frame and resolve a hydrofoil wake. After zero-shot transfer to a real water tunnel, the agent locates wakes of randomly placed hydrofoils and measures them at higher resolution and lower flow-statistics error than a fixed view, in roughly half the time of a full scan and with far fewer image pairs.
Interpretation
The measurement camera's position changes from something an operator sets before the run into something the agent decides during the run from the image pairs already acquired. Conventional PIV fixes magnification, camera distance and inter-frame time before the experiment under established design rules; robot-mounted PIV traverses a pre-planned path and stitches afterwards, and the acquired data likewise does not feed back into camera positioning. In the real water tunnel, across two hydrofoils absent from training, a NACA 0015 and a NASA SC(2)-0714, at six location-angle combinations each, 41 of 42 episodes settled on a view containing the whole wake, whereas the fixed wide view covering about 260 mm of the test section contained the whole wake in 7 of 11 settings.
The agent sees the scene only through a velocity field and its uncertainty, and moves the camera to the pose that captures the whole wake at the highest magnification. It does not look at particle images directly; it receives the displacement field and uncertainty that LIMA, a CNN-based PIV estimator, infers from four frame pairs, and uses the uncertainty to separate the wake from unlit shadow regions. The policy is trained entirely in NVIDIA Isaac Sim on 719 cases from 372 hydrofoil configurations at two free-stream speeds, with wakes from CFD, and is deployed on hardware without fine-tuning or real-world data, capturing the wake within a median of two steps from a random initial view.
Closer framing puts more pixels on the same wake and improves recovery of the flow statistics. The fixed wide view images the sheet at a pixel size of about 0.136 mm, while the agent's settled views have a median pixel size of 0.090 mm, roughly two thirds of that. Across the eleven hydrofoil settings, relative to a 15-tile stitched scan, the agent's views reduce both the relative summed error and the signed bias of three flow-field variables in the wake region, for example lowering one variable's error from 0.419±0.212 to 0.205±0.049.
Perspective
The demonstrated scope is wake measurement for two 3D-printed hydrofoils in a small laboratory water tunnel, with the agent re-locating the wake from random initial views and after a mid-episode hydrofoil change; the authors expect transfer to other wind and water tunnels, conditional on accurate CAD models, adequately represented lighting in simulation and extensive camera calibration.
The agent does not adjust imaging parameters such as exposure time, inter-frame time and aperture, which also affect flow-reconstruction quality; extending the policy to control these parameters is named as future work. Setting N15-3, where the camera dropped frames during the mosaic scan, is shown but excluded from comparisons with the mosaic.