The catalogue
Experiments.
Everything we run, measure, or simulate — coupled crowd↔CSI simulations, real BLE/CSI captures, notebook analyses, and hybrid co-simulations. Each experiment links to the campaigns that executed it, the sessions and runs behind it, and the building it lives in.
18 of 18 experiments
| Experiment | Kind | Domain | Execution | Campaigns | Sessions | Runs | Last activity | Evidence |
|---|---|---|---|---|---|---|---|---|
| Body-as-scatterer EM fidelity — what the dielectric cylinder hides (ray-traced) A single homogeneous dielectric cylinder + one calibrated per-body loss is an adequate scatterer primitive for amplitude-bundle statistics IF the per-body loss and radius are jointly identifiable from the attenuation *d… | simulation | csi | planned | 1 | — | — | — | map ↗ |
| CSI amplitude under static occupancy — the blockage↔variance crossover (ray-traced) As occupancy rises, mean CSI amplitude falls and saturates (L4.6 LOS blockage) while the amplitude bundle's coefficient of variation and Rician-K spread rise (L4.5 multipath); the two signatures cross at an occupancy th… | simulation | csi | executed | 1 | — | 16 | — | map ↗ |
| Does CSI occupancy-sensitivity generalize across floorplans? (ResPlan ensemble) The CV-vs-occupancy relationship (the L4.5 variance feature) is geometry-dependent: its slope and offset vary materially across diverse real apartment layouts. | simulation | csi | planned | 1 | — | — | — | map ↗ |
| Locating the blockage-dominant regime — through-wall, sparse-LOS CSI When the Tx–Rx link is wall-obstructed and LOS is sparse (Tx and Rx in different rooms), the mean-attenuation signature (L4.6) leads at low occupancy and the blockage↔variance crossover sits at N*>0 — the Depatla throug… | simulation | csi | planned | 1 | — | — | — | map ↗ |
| Time-resolved CSI under real crowd motion (coupled JuPedSim → Sionna) Motion-driven *temporal* CSI variance — the dynamic-to-noise / bundle-width feature Choi/Ling actually exploit — emerges only under moving crowds, not the static placements of csi-static-occlusion. | simulation | csi | executed | 1 | — | 6 | — | map ↗ |
| Active Injection vs Passive Sniffing Non-uniform (bursty) passive packet arrival destroys Doppler/activity features even when the MEAN packet rate comfortably exceeds the Nyquist bound, because uneven sampling spreads the motion signature into spectral lea… | unclassified | — | idea | — | — | — | — | — |
| BLE-Anchored CSI Ground Truth — validating the mobile-app counting chain against a small in-the-loop campaign Counting unique BLE advertisements from the gamification mobile app, observed via [[_knowledge/hardware/esp32-c6|ESP32-C6]] listeners, matches a prescribed walk-task schedule and a spot-check manual count within ± 1 per… | unclassified | — | idea | — | — | — | — | — |
| BLE-Assisted CSI Ground Truth Collection BLE advertisement counting from a mobile app provides reliable automated ground truth for CSI crowd counting experiments, eliminating the need for manual head counts. | unclassified | — | idea | — | — | — | — | — |
| CSI Drift and Periodic BLE Calibration — the thesis sawtooth (a) A CSI crowd-counting model frozen at Day 0 measurably degrades over 30 days as the environment changes, and (b) short trigger-based BLE calibration campaigns (10–30 min, mobile-app-driven) restore the model to withi… | unclassified | — | idea | — | — | — | — | — |
| CSI Model Drift Measurement Over Time CSI-based crowd counting accuracy degrades significantly over 7-30 days without recalibration, driven by environmental changes (furniture, temperature, door state), with degradation rate correlated to the magnitude of e… | unclassified | — | idea | — | — | — | — | — |
| Hybrid BLE+CSI Occupancy Estimation Fusing BLE advertisement counting (device-based) with CSI sensing (device-free) produces more robust occupancy estimates than either modality alone, especially in scenarios with partial phone penetration. | unclassified | — | idea | — | — | — | — | — |
| Hybrid Fusion and Cross-Geometry Transfer — BLE+CSI fusion across phone penetration and across rooms A BLE+CSI fusion model trained in room A (a) achieves lower MAE than either modality alone across all five phone-penetration scenarios (S1 100 % → S5 0 %), AND (b) transfers to room B via a single 30-minute BLE-anchored… | unclassified | — | idea | — | — | — | — | — |
| Multi-Room Generalization with BLE Transfer A CSI crowd counting model trained in Room A can be transferred to Room B using only a 10-30 minute BLE calibration campaign, achieving >80% of the accuracy of a model fully trained in Room B. | unclassified | — | idea | — | — | — | — | — |
| Packet-Rate Floor for Motion Sensing The detectability of walking-speed motion in CSI collapses once the packet reception rate (PRR) falls below the Nyquist bound for that motion's Doppler shift; the empirical breakpoint in a downstream motion/occupancy-ch… | unclassified | — | idea | — | — | — | — | — |
| Periodic BLE Calibration Campaigns Short BLE-assisted calibration campaigns (10-30 minutes) can restore CSI crowd counting accuracy to within 5% of the original Day 0 performance, at a fraction of the cost of full retraining. | unclassified | — | idea | — | — | — | — | — |
| Sensing Node Bring-Up and Distributed Plane — ESP32 + Pi 5 + AX210 + mobile-app gamification loop The lab inventory (10× ESP32-C5, 10× ESP32-C6, 5× ESP32-S3, 14× Pi 5 + M.2 HAT + AX210, mobile app) can be brought up as a single distributed sensing plane — synchronised to sub-100 ms across all nodes, end-to-end groun… | unclassified | — | idea | — | — | — | — | — |
| Simulation Sandbox — JuPedSim + BLE + CSI-proxy coupled, mechanism + harness validation A pure-Python coupled simulation of pedestrian dynamics (JuPedSim), BLE RSSI (log-distance + log-normal), and a statistical CSI proxy is sufficient to (i) validate the data-plane harness that EXP-P1 and EXP-F* will reus… | unclassified | — | idea | — | — | — | — | — |
| Synthetic Continuity-Residual Fusion Sandbox — JuPedSim + Sionna + BLE Co-Simulation Adding the continuity-residual penalty $\lambda \| \partial \rho/\partial t + \nabla \cdot (\rho \mathbf{v}) \|_2^2$ to the supervised loss reduces per-cell occupancy MAE across **every** model architecture in a 9-model… | unclassified | — | idea | — | — | — | — | — |
Campaigns.
A campaign is a written brief — a question, success criteria, a budget — that an autonomous supervisor executes as one or more sessions. Planned campaigns have a brief but no sessions yet.