Multi-Room Generalization with BLE Transfer

Current assessment (2026-09-15)

Superseded design retained for provenance, reviewed September 15, 2026. No primary data are attributed to this card. The six-floor simulated transfer estimate remains uncertain, and phone observations need an identified person scale. The original room-transfer target is a design threshold, not a delivered result.

The six-floor simulated transfer estimate remains uncertain, and phone observations need an identified person scale. The original room-transfer target is a design threshold, not a delivered result.

Next evidence needed: Use EXP-F3 for the current protocol and evidence. The earlier plan below is not an instruction to reopen this experiment.

Basis: Thesis Narrative, the September 10 synthesis, and the September 15 register review. See the review record for evidence sources and outstanding checks.

Earlier protocol and dated results

The methods and observations below retain their original scope. Read their readiness statements and proposed next steps as historical. The current assessment above records what has changed; frozen thresholds have not been retrospectively revised.

The problem, in plain words

Imagine learning to judge a crowd's size by the echo of your own voice in a room. In a small carpeted meeting room the echo is short and muffled; you learn "loud short echo ≈ 5 people." Walk into a big empty seminar hall and the same 5 people produce a long, ringing echo — your rule now reads far too high. You did not get worse at counting; the room changed the physics of the measurement.

Wi-Fi CSI crowd counting has exactly this disease. The model learns a mapping from radio-channel patterns to headcount, but those patterns depend on the room's size, its furniture, and its wall materials. Move to a new room and the mapping silently breaks. santos2024_1e39 ↗ documents this "lack of environment independence" as a core unsolved gap; wang2026_2758 ↗ and chen2023_5cbd ↗ frame it as the barrier to scaling wireless sensing.

The expensive fix is to collect a fresh training set — hours of labelled occupancy — in every new room. The idea here is to replace that with a short BLE calibration campaign: a handful of participants carry phones broadcasting Bluetooth beacons for 10–30 minutes, which gives cheap, automatic ground-truth headcounts to fine-tune the existing model onto the new room. If it works, deploying to a new room costs minutes, not hours.

What we are trying to prove

• Hypothesis (falsifiable): a model trained in Room A, fine-tuned only on a 30-minute BLE-anchored campaign in Room B, reaches ≥ 80 % of the accuracy of a model fully trained on hours of Room B data — and a 10-minute campaign already recovers > 50 % of the zero-shot-to-native accuracy gap.
• What a null means: if the short campaign does not close the gap — if fine-tuning on 10–30 minutes lands no better than zero-shot transfer, or if it overfits the tiny target set and does worse — then BLE calibration is a time-drift tool, not a geometry-transfer tool, and cross-room deployment needs a different mechanism (domain-invariant features, or simply more target data). The adjacent c-feature-type-transfer result makes this null a live outcome, not a formality.
• The sharper question underneath: is the thing that transfers a better-calibrated fingerprint (what BLE fine-tuning gives you) or a different feature family (kinematic features that were geometry-robust all along)? EXP-004 as written tests the first; the simulation evidence argues the second may dominate.

How the experiment works (plain method)

Two meaningfully different rooms are chosen to maximise the domain gap — e.g. a small dense meeting room (Room A) versus a large sparse seminar room (Room B), different TX–RX distances, different multipath.

1. Full training in Room A — collect ~2 h of CSI + BLE ground truth, train three model families (Random Forest, LSTM, CNN). This is the source model and the Room-A upper bound.
2. Zero-shot transfer to Room B — apply the Room-A model to Room B with no retraining. Expect a large accuracy drop; this is the problem evidence and the lower bound.
3. BLE-assisted transfer — run a 10-minute BLE campaign in Room B (≈5 participants), fine-tune the Room-A model on it, evaluate on held-out Room-B data. Repeat at 20 and 30 minutes to trace the duration → accuracy curve.
4. Native Room-B baseline — train a fresh model on ~2 h of Room-B data. This is the gold standard the transfer is scored against.
5. (optional) Bidirectional check — is A→B as easy as B→A, or is transfer asymmetric?

The headline metric is transfer efficiency = BLE-transfer accuracy ÷ native Room-B accuracy (target > 0.8), and the data-efficiency multiplier (10 min vs 120 min ≈ 12× less labelled data).

The plan status — and what adjacent sims already tell us

EXP-004 itself: zero runs. No hardware campaign, no BLE-calibrated-transfer arm, in either this card or its EXP-F3 successor (EXP-F3 is status: planned). Every number in the original "expected accuracy" table is a placeholder, not a result. Treat the >80 % target as an aspiration to be tested, not a finding.

What we can honestly borrow from the simulation platform, because it bears directly on the premise:

The takeaway for the design: step 2 (zero-shot fails) is essentially pre-confirmed in-silico. The open, untested question is whether step 3 (short BLE fine-tune) actually closes the gap — and whether it beats the cheaper alternative of simply switching to geometry-robust kinematic features.

Key references

• santos2024_1e39 ↗ — the environment-dependence problem statement this experiment attacks.
• wang2026_2758 ↗, chen2023_5cbd ↗ — cross-domain / generalizability taxonomies that frame the transfer gap.
• hou2022_2a52 ↗, hou2023_bf83 ↗ — few-shot cross-domain counting; the identifiability basis for a short campaign.
• khan2023_b7c5 ↗, zhang2018_138d ↗, zhang2025_a250 ↗ — structured transfer / domain-adaptation baselines the naive-fine-tune arm must beat.
• demrozi2021_bf55 ↗ — the BLE calibration modality's real-world basis.
• zou2018_1590 ↗ — device-free CSI counting baseline.
• zhang2026_ccac ↗ — the reproducibility bar the SWE voice invokes against the phantom scaleout session.
• huang2025_060d ↗ — a public real-CSI anchor for the red-team's cross-check demand.