| # Evaluation Card |
|
|
| ## Benchmark Task |
|
|
| Given a temporal history of six body-fixed induced-flow sensors, predict the leader-centered relative state of a nearby fish. |
|
|
| ## Inputs |
|
|
| Inputs are sparse velocity histories from processed HDF5 arrays. The main feature sets are: |
|
|
| - `raw`: raw per-sensor induced velocity features. |
| - `raw_norm`: raw features plus amplitude-normalized velocity features and the local scale. |
|
|
| ## Targets |
|
|
| Location-only runs predict `(delta_x, delta_y)`. Full-state runs predict `(delta_x, delta_y, theta_rel)`. |
|
|
| ## Metrics |
|
|
| Position error is reported as `sqrt(RMSE_delta_x^2 + RMSE_delta_y^2)` in body lengths. Relative-orientation error is circular mean absolute error for `theta_rel`, reported in degrees. |
|
|
| ## Pose-Holdout Split |
|
|
| All wake phases for each held-out test pose are assigned to the test partition. Training and validation are drawn from the remaining non-test samples, so training and validation may contain different phases of the same non-test pose. This matches the experiment driver and prevents repeated-pose leakage into the test set. |
|
|
| ## Main Ablations |
|
|
| - History length. |
| - Sensor ablation. |
| - Flow component: wake-only, potential-only, wake-plus-potential. |
| - Common raw-noise stress test. |
| - Architecture screen. |
| - Sample-size convergence. |
| - Self-signal controls. |
|
|
| ## Noise Protocol |
|
|
| Noise is added to raw `u/v` channels before feature construction and standardization. The component/noise sweep uses a common absolute raw-sensor noise scale across flow components. |
|
|
| ## Self-Signal Controls |
|
|
| Aligned and randomized-orientation datasets separate external-only, self-only, and total-like external-plus-self signals. Self-only is a negative control for leakage. |
|
|
| ## Expected Manuscript Numbers |
|
|
| Stored CSVs in `data/results/` contain the manuscript numbers. The main full-state close-wake result is approximately `0.040 L` position RMSE and `0.53 deg` `theta_rel` MAE. Randomized-orientation external-plus-self full-state control is approximately `0.087 L` close-wake position RMSE and `1.08 deg` close-wake `theta_rel` MAE. |
|
|
| ## Failure Modes and Limitations |
|
|
| Potential-only inputs can be accurate without noise but fail under the common absolute noise stress test. Farther wake regions are harder than close-wake regions. The artifact does not evaluate full biological pressure sensing, multi-neighbor assignment, or closed-loop control. |
|
|
