ART of PIV: Agentic Real-Time Particle Image Velocimetry for Active Flow Control
Organizations: Institute for Dynamic Systems and Control, ETH Zurich, 8092 Zurich, Switzerland
Abstract
Particle image velocimetry (PIV) measures fluid motion from images of tracer particles, but its accuracy depends on choices throughout the estimation pipeline, including image preprocessing, flow estimation, and postprocessing. These choices are typically left to expert tuning, and a fixed configuration can become less accurate as conditions change. This matters especially when PIV supplies feedback for real-time control, where measurements must be both accurate and fast, and flow conditions vary. We introduce ART of PIV -- short for Agentic Real-Time and also a nod to the art of tuning a PIV pipeline -- which automates pipeline choice for each image pair. A learned agent selects a configuration according to a chosen measurement objective from a library assembled for a specified speed budget. Across cross-correlation, classical optical-flow, and learning-based estimators, ART of PIV reduces average end-point error (AEPE) by 12.2-22.4 % relative to the best single estimator at matched speed, with gains extending to 1000-4000 image pair/s. In a simulated local velocity regulation task, it reduces mean normalized closed-loop cost by 13.24-13.59 percentage points relative to a tuned fixed estimator, with the same classical controller in both cases. In a tabletop water channel, we illustrate two ways to tune ART of PIV for a fluidic pinball of three rotating cylinders: selecting between complementary optical-flow tunings for accuracy, and selecting regions of interest with coherent rotation or shear using a common flow estimator. For the first variant, we train a DreamerV3 reinforcement learning controller directly in the physical world and demonstrate tracking of a prescribed flow field in the wake of a fluidic pinball.
Figures & tables
| AEPE (px) | |||||
| Estimator | Library spans | Speed ( image pair/s ) | Best single | ART of PIV | rAEPE (%) |
| Learned | |||||
| RAFT32-PIV | flow classes | 1.7 | 0.0120 | 0.0096 | 20.3 |
| Classical optical flow | |||||
| DIS | hyperparameters | 20 | 0.1498 | 0.1268 | 15.3 |
| DIS | hyperparameters | 1000 | 0.1577 | 0.1385 | 12.2 |
| Feedback | (pp) | (pp) | Successful episodes |
|---|---|---|---|
| Fixed specialist 0 | 0/30 | ||
| Fixed specialist 1 | 11/30 | ||
| Reference | 18/30 | ||
| ART of PIV | 30/30 | ||
| Whole-frame oracle | 30/30 | ||
| Ground-truth feedback | 30/30 |
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
| # models | Baseline (px) | Oracle (px) | Classification (px) | Regression (px) | Bandit (px) |
|---|---|---|---|---|---|
| 1 | 0.1370 | 0.1370 | — | — | — |
| 2 | 0.1370 | 0.1196 | 0.1208 | 0.1211 | 0.1219 |
| 3 | 0.1370 | 0.1073 | 0.1086 | 0.1090 | 0.1096 |
| 4 | 0.1370 | 0.1064 | 0.1081 | 0.1081 | 0.1101 |
| 5 | 0.1370 | 0.1056 | 0.1079 | 0.1080 | 0.1102 |
| 10 | 0.1370 | 0.1041 | 0.1075 | 0.1068 | 0.1101 |
| Regression target | Reference | ||||
|---|---|---|---|---|---|
| None (raw EPE ) | Per-pair mean | Log | Classification | Oracle | |
| AEPE (px) | 0.1084 | 0.1072 | 0.1067 | 0.1076 | 0.1034 |
| Estimator | Speed ( image pair/s ) | Agent | AEPE (px) | rAEPE (%) | Agent time (ms) |
|---|---|---|---|---|---|
| DIS | 1000 | deployed | 0.1279 | 15.1 | 0.16 |
| larger | 0.1272 | 15.6 | 0.27 | ||
| OpenPIV | 1000 | deployed | 0.1827 | 21.0 | 0.11 |
| larger | 0.1823 | 21.2 | 0.21 | ||
| OpenPIV | 4000 | deployed | 0.1828 | 21.0 | 0.11 |
| larger | 0.1824 | 21.2 | 0.21 |
| Estimator | Speed ( image pair/s ) | Method | Mean | Std | Median | p90 | p95 | p99 | Max |
|---|---|---|---|---|---|---|---|---|---|
| RAFT32-PIV | 1.7 | Best single | 0.0120 | 0.0090 | 0.0103 | 0.0200 | 0.0270 | 0.0451 | 0.0790 |
| ART of PIV | 0.0096 | 0.0073 | 0.0077 | 0.0179 | 0.0227 | 0.0365 | 0.0424 | ||
| DIS | 20 | Best single | 0.1498 | 0.1270 | 0.1055 | 0.2767 | 0.3793 | 0.6275 | 1.4324 |
| ART of PIV | 0.1268 | 0.0967 | 0.0944 | 0.2384 | 0.3210 | 0.5130 | 0.6733 | ||
| DIS | 1000 | Best single | 0.1577 | 0.1219 | 0.1133 | 0.3016 | 0.3882 | 0.6307 | 1.0053 |
| ART of PIV | 0.1385 | 0.1058 | 0.1023 | 0.2648 | 0.3499 | 0.5973 | 0.7854 |
| Estimator | Speed ( image pair/s ) | Better | Tie | Worse |
|---|---|---|---|---|
| RAFT32-PIV | 1.7 | 98.1 | 0.0 | 1.9 |
| DIS | 20 | 68.9 | 13.7 | 17.4 |
| DIS | 1000 | 78.5 | 0.2 | 21.3 |
| OpenPIV | 1000 | 65.6 | 21.9 | 12.4 |
| OpenPIV | 4000 | 61.5 | 27.3 | 11.2 |
| Component | Symbol or quantity | Nominal value |
|---|---|---|
| Fluid and forces | Viscosity | |
| Width ; dimensionless wavenumber ; coupling | ; ; | |
| Initial condition | ; | ; |
| Episode variation | Independent uniform draws: amplitudes , viscosity | |
| Target and work area | ; side of | ; |
| Work-area center | , |
| Estimator | Patch (px) | Stride (px) | Search iterations | Pyramid levels | Finest level |
|---|---|---|---|---|---|
| Specialist 0 | 15 | 7 | 16 | 4 | 1 |
| Specialist 1 | 9 | 8 | 12 | 4 | 0 |
| Reference | 13 | 10 | 8 | 5 | 0 |
| Feedback | (pp) | (pp) | Successful episodes |
|---|---|---|---|
| No control | 0/30 | ||
| Fixed specialist 0 | 0/30 | ||
| Fixed specialist 1 | 7/30 | ||
| Reference | 15/30 | ||
| Whole-frame oracle | 29/30 | ||
| Window oracle | 30/30 |
| Parameter | Nominal | Tested values | (pp) | (pp) |
|---|---|---|---|---|
| Fluid and initial condition | ||||
| ( ) | ||||
| ( ) | ||||
| Parameter | Nominal | Tested values | (pp) | (pp) |
|---|---|---|---|---|
| Observation and imaging | ||||
| Work-area side for feedback (px) | ||||
| Work-area center for feedback | ||||
| (s) | ||||
| Particle density (particles/px) | ||||
| Spot diameter (px) | ||||
| Parameter | Nominal | Tested values | (pp) | (pp) |
|---|---|---|---|---|
| ( ) | ||||
| (s) | ||||
| (s) |
| Step | Operation | Parameters |
|---|---|---|
| 1 | intensity_capping | |
| 2 | stretch_contrast | |
| 3 | Spatial reduction | , antialiased linear |
| 4 | Temporal averaging | 3 fields at 30 Hz |
| 5 | Observation scaling | , clipped |
| 6 | Controller input | , CNN at 30 Hz |
| Configuration | Finest level | Levels | Patch size (px) | Patch stride (px) | Iterations |
|---|---|---|---|---|---|
| A | 0 | 1 | 19 | 11 | 4 |
| B | 1 | 3 | 13 | 9 | 4 |
| C | 0 | 2 | 13 | 7 | 6 |
| Variant | Label | Cylinder 1 (RPM) | Cylinder 2 (RPM) | Cylinder 3 (RPM) |
|---|---|---|---|---|
| Tuning | A | 150 | -150 | -150 |
| Tuning | B | 0 | 150 | -150 |
| Region | R1 | 150 | 150 | 150 |
| Region | R2 | 150 | 150 | -150 |
| Region | R3 | 150 | -150 | 150 |
| Region | Left column (px) | Top row (px) |
|---|---|---|
| R1 | 256 | 64 |
| R2 | 0 | 256 |
| R3 | 512 | 384 |
| Library | Our choice | Correct/total | Agreement (%) |
|---|---|---|---|
| Tuning | A | 1 799/1 799 | 100.00 |
| Tuning | B | 1 799/1 799 | 100.00 |
| Region | R1 | 1 795/1 799 | 99.78 |
| Region | R2 | 1 799/1 799 | 100.00 |
| Region | R3 | 1 799/1 799 | 100.00 |
| Setting | Configuration | Native | ||
|---|---|---|---|---|
| A | A | 0.3666 | 0.3248 | 0.4609 |
| A | B | 0.4585 | 0.4511 | 0.4861 |
| B | A | 0.6170 | 0.6095 | 0.9951 |
| B | B | 0.4864 | 0.4765 | 0.5359 |
| Training | Completed seeds | Test agreement (%) |
|---|---|---|
| With augmentation | 5/5 | |
| Without augmentation | 5/5 |
Algorithm Selection
Algorithm selection predicts which solver, model, or strategy will work best for a given instance, context, or computational budget.
See all 54 papers