Asking the fleet what it is doing…
monad-knowledge Wi-Fi sensing lab · FIIT STU
Campaign

CSI x BLE fusion under motion — does periodic BLE recalibration bound the cross-environment drift of temporal-CSI counting?

c-csi-ble-fusion · exp-csi-crowd

Archive snapshot, as of 18 h ago — the run corpus is rebuilt once a day, so this page is not a live reading. The fleet panel is the live one; it refreshes every 30 s.

Sessions

state
Session State Runs Synthesis Criticism Figures Verdict
MYRQP09E 2026-06-25T21:06 finished 3 ✓ ✓ — POSITIVE: CSI was under-featured, not irredeemably weak. Turning Doppler micro-fading ON + using a richer intra-frame-fading feature lifts CSI->count from R2=0.09 (quasi-static, hand amplitude-CV) to R2=0.36, nMAE 0.19 on the 10-link dense layout at varying high occupancy (occ->18) — ~4x the count variance explained. Doppler alone lifts hand-CV to R2=0.20; the rich per-link intra-frame stats (linear ridge) take it to 0.36. The neural learned feature OVERFITS at this data scale (3 runs, ~380 train samples -> R2=-0.43 even tamed) — so the LINEAR rich-feature model is the right complexity now; a genuine learned model needs much more coupled data. The lever is temporal CSI structure (Doppler) + a richer feature, NOT more anchors. This lifts the staleness-switch's CSI-side floor and makes CSI a materially more useful fusion complement.
DRM17SVP 2026-06-25T20:56 finished 6 ✓ ✓ — Denser-CSI + higher-occupancy did NOT meaningfully lift the CSI floor. On resplan-12439 with a NEW 10-Rx 'dense10' experiment + varying high occupancy (occ to 18, real arrival/departure variance), the CSI temporal-CV explains almost no count variance: 10-link per-link-ridge R2=0.09 (the only config above 0), 10-link scalar 0.00, 5-link both ~0/negative; nMAE ~0.23-0.26. Both levers help DIRECTIONALLY (10>5 links, per-link>scalar) but the absolute signal stays very weak. The earlier 0.16 was inflated by near-constant (seated-stay) occupancy. Conclusion: more links/occupancy is NOT the lever — the quasi-static ray-traced amplitude-CV feature is a fundamentally weak count predictor here. The real remaining lever is richer TEMPORAL CSI features (doppler micro-fading, off in these runs for frame alignment; or a learned spectro-temporal feature), or real CSI. Reinforces the arc: BLE device-counting is the load-bearing modality; CSI is a weak coarse fallback in the staleness-switch.
NKDDWCH0 2026-06-25T20:31 finished 9 ✓ ✓ — Staleness-switch fusion (w=exp(-staleness/tau), tau~16s) is validated as the right architecture: ONE estimator that tracks the lower envelope of {BLE-held, CSI} across the full BLE-cadence range — matches/edges BLE when fresh (0.73 dense), falls back to the CSI floor when the anchor goes stale (1.6 vs held-BLE 2.8 at 80s), and beats the naive offset-hold blend throughout. It does not strictly dominate (stays within ~0.1 person of the best modality at every cadence) and beats each modality in the other's regime. HONEST LIMIT: the per-floor BLE-adapted CSI map gave ~0 gain (cold 1.50 ~= adapted 1.51) — the CSI count signal is the weak link at these low crowd sizes / 3-link layouts, not the calibration; so the switch's sparse-regime floor is capped at CSI's ~1.5. Fusion ceiling = CSI quality, which needs more links / higher density / a better feature, not better calibration.
YG02RX48 2026-06-25T20:25 finished 9 ✓ ✓ — Sparse-BLE-cadence axis: the naive always-anchor fusion is DOMINATED at both ends. Dense BLE (<~7s): held-BLE ~= fused ~= 0.77 (CSI adds nothing, confirms v2). Sparse BLE (>~12s): plain CSI-only (1.39, flat) beats both held-BLE (up to 2.68) and fused (up to 2.01). The fused only beats held-BLE past ~25s (CSI shape vs very-stale BLE) but never beats the BEST-of-{BLE,CSI}. Design lesson: the right architecture is a CONFIDENCE-WEIGHTED SWITCH (trust BLE while fresh, fall back to CSI when the anchor goes stale), not a blind blend. CSI's distinct value is as the stale-BLE fallback, not an always-on improvement.
Z5ZAKF7E 2026-06-25T20:11 finished 9 ✓ ✓ — Fixed result (v2): with the v1 confounds removed — BLE modelled as an independent device-counter (noisy occupancy, not an RSSI->count fit on target truth), matched through-traffic occupancy across floors, 3 seeds — periodic BLE recalibration still bounds the CSI source-map geometry drift: cross-floor MAE 1.39 -> 0.82 persons (41%, error bars hold). The honest new nuance: fused (0.82) ~ BLE-only (0.77), so at these crowd sizes with a dense BLE counter, CSI adds little over just using BLE; the fusion's value is the high-rate shape BETWEEN sparse BLE ticks, which a static MAE underweights.
336Q16PW 2026-06-25T19:35 finished 6 ✓ ✓ — Mechanism shown in-silico: a target-side absolute anchor can bound the drift a relative temporal-CSI count feature suffers across floors. The 3.5->1.0-person ("70.7%") figure is NOT a clean drift-recovery rate — it is confounded by (a) the BLE anchor being fit to the target's own ground truth (leakage) and (b) source(occ<=10)/target(occ<=6) occupancy-range extrapolation. Hybrid-Fusion first in-silico empirical; magnitude unproven.

Brief

Question

c-csi-crowd-temporal proved the simulator fusion (walk-notebook -> Sionna, per-frame ray-traced CSI under real motion) but explicitly deferred "BLE coupling (a later fusion experiment)." This is that experiment — and the data fusion half of the thesis. With ble.enabled, one solved channel per frame yields BOTH a temporal CSI tensor and a co-registered BLE-RSSI stream. The temporal-CSI counting feature (CV over a sliding window — ling2024 EasyCount, choi2022 Wi-CaL) is a high-rate relative density signal whose calibration does not transfer across geometry (santos2024; the c-csi-cross-geometry-resplan x3.8-x7.7 inflation). The BLE anchor is a coarse, noisy, but geometry-robust absolute reference. Does fusing them — CSI for the shape, periodic BLE for the scale — bound the drift that sinks CSI-only?

Framing discipline. In-silico, self-authored crowd + synthetic CSI/BLE with a self-authored Gaussian BLE noise. Accuracy against that noise is circular; the honest result is the mechanism: does periodic BLE recalibration bound cross-environment drift? Motivation for the Hybrid-Fusion chain, not a hardware claim — the field defeater is IP-106.

Why now — the thread this closes

The thesis Hybrid-Fusion chapter (weight 0.6) has zero supporting empirics, yet ble-periodic-calibration, recalibration-trigger-from-drift and ble-ground-truth-sufficiency all assert that a BLE anchor corrects CSI drift. The coupled chain + co-registered BLE is exactly the testbed: it produces the two modalities from one channel, on the same floors where CSI-only is already known to fail cross-geometry. The cross-environment LOEO real-data finding (recalibration recovers ~42% of the lost accuracy) is the field counterpart this in-silico run motivates.

Datasets / fidelity exercised

  • ResPlan floors with placed csi-link-* experiments: resplan-12439 (furnished, 16 seats — a seated-wave + movers crowd, the richest occupancy(t) shape) as the calibration source; resplan-1374, resplan-16157 (through-traffic waves) as transfer targets.
  • Co-registered BLE (BleParams, same CFR) and sub-frame Doppler micro-fading (DopplerParams, 2026-06-25) so the 2.4 vs 5.0 GHz bands actually differ in temporal CV (without it the quasi-static RT is band-independent — the c-csi-band-calibration finding).

What the supervisor does

  1. Fan out exp-csi-crowd over {3 floors} x {2.4, 5.0 GHz}, ble.enabled, doppler.enabled, time-varying crowd, trajectory_frame_stride: 24 (≈44 s/run on CPU). Each floor uses its own fusion_experiments[floor] device layout.
  2. Run fusion_count (runner python, IP-087): per (floor, band) compute occupancy(t), CSI-CV, BLE-attenuation; the CSI-only / BLE-anchored / fused estimators; the per-floor timeline; and the cross-floor transfer (source map -> targets, CSI-only vs BLE-recalibrated-fused).
  3. Synthesise the drift-recovery (transfer-MAE reduction) with the in-silico framing.

Figure render request

fusion_count — (a) per-floor fused timeline (truth vs CSI-only-drift vs fused, with BLE recal ticks); (b) cross-floor transfer MAE bars (CSI-only vs fused) per target floor; (c) the band contrast (2.4 vs 5.0 GHz temporal-CV).

Out of scope

  • The continuity-equation prior (continuity-prior-regularizes-density) — the physics-fusion PINN is a separate campaign needing the python-ml runner.
  • Real-channel transfer — the synthetic-to-real gap (synthetic-csi-sim-to-real-transfer) and the hardware noise model are IP-106 / exp-csi-calibration.
  • Estimator sophistication — affine recalibration is deliberate (interpretable, non-circular); a learned fusion head is a follow-on once the mechanism is shown.

Expected interpretation

  1. Fused transfer-MAE << CSI-only transfer-MAE -> "periodic BLE recalibration bounds the cross-environment drift of temporal-CSI counting." Upgrades ble-periodic-calibration toward supported (in-silico) and gives the Hybrid-Fusion chapter its first empirical.
  2. No improvement -> the BLE anchor is too coarse/saturated to recalibrate at these crowd sizes (cf. the c-ble-csi-coregistration saturation) — a clean negative that pushes fusion toward a multi-anchor or CSI-feature fix rather than relocation.
  3. Chain/co-registration fails -> a platform finding on the coupled BLE path.