sailswarm-fusion-scorer-v1b

SailSwarm fusion scorer v1b — the live per-sector obstacle probability (gradient-boosted)

The scorer the SailSwarm obstacle-detection box runs on the boat (University of Konstanz, Lake Constance; fisheye RGB + FLIR Lepton 3 thermal + TI AWR1843 mmWave radar on a Raspberry Pi 4). For each 15° bearing sector it emits a calibrated probability that the sector holds a navigation-relevant obstacle. It replaced the gated-mixture v1a (NexusDwin/sailswarm-fusion-scorer-v1a) on 2026-09-02 after the first audited low-light truth showed the interactions matter: AP 0.883 vs 0.731 (incumbent n/3 vote) vs 0.639 (v1a) on the audited night holdout.

Three bundles, one per deployment mode of the box (deploy/shadow_mode.sh):

Incumbent (rule-based n_sensors/3) on the same holdout: AP 0.673, ECE 0.212. Full tables incl. day/dusk strata and the night set: reports/.

Calibration. The isotonic calibrator inside each bundle is fitted out of fold (train.py --oof-calibration): the train-fit calibrator that shipped on 2026-09-02 read 0.803 AP / ECE 0.32 on the audited holdout against 0.886 / 0.14 out of fold. Consume predict_proba_calibrated.

Known limit (measured on the 2026-09-08 lake trials). With no radar return the scorer floors camera-only evidence at p ≈ 0.24: day marginal weights per sensor hit fisheye +0.02 / thermal +0.02 / radar +0.23; inside the night trial window every sensor's marginal weight is ≤ +0.04. The boat-side policy therefore takes the per-sector max of this p and the fused vote (--evidence max). The next retrain adds the audited night frames of that session.

Loading

import joblib                      # scikit-learn == 1.9.0 (the pin on the box)
bundle = joblib.load("fusion_scorer_live_v1b_oofcal.joblib")   # GBTScorerBundle, see models.py
p = bundle.predict_proba_calibrated(X)   # X: the bundle's own column list, bundle.columns

models.py and fusion_model.yaml are the exact versions the bundles were built with (repo SailSwarm-ObstacleDetection, scripts/fusion_model/). Features come from build_features.py (schema v4), targets from audited frames in bearing space.

reports/night_retrain_2026-09-12 (added 2026-09-15)

The session-disjoint evaluation reported in the ICRA 2027 submission (its Table II) and the night retrain that followed the 2026-09-08 lake trials. Models were trained on the six earlier sessions with luminance-gated targets; the whole 2026-09-08 session is held out (day 532 sector rows, twilight 917, night 1,561, of which 100 night boxes were drawn on the thermal image). Source of the paper's numbers: ablation_day_v3d_0908day.{md,json} (rows incumbent = rule-based vote, v1a_gate, full_oofcal = GBT with context, rgb_only, thermal_only, radar_only; the day/dusk/night AP columns are the table).

Headline: at true night the rule-based vote (AP 0.582) beats every learned variant (at most 0.520), while thermal-only and radar-only GBTs recall 0.90 and 1.00 of the audited persons against 0.17 for the vote. Leave-one-night-group-out cross-validation (bakeoff_nightcv3d_*) preserves the ordering. retrain_report_ab_v3d.md is the matched replay of the retrained bundle against the live one on the trial evening: the retrained model raises the daytime base probability to about 0.56, which would have turned 15 of the 20 closed-loop decisions into "blocked" under the 0.5 policy threshold, so the live bundles in this repository are unchanged. Verdict: data-limited, not model-limited; the binding need is audited night frames with obstacles inside the radar's 9 m envelope.