m monad-knowledge FIIT STU · open research vault
Campaign

Coupled crowd → CSI on a ResPlan floor — synthetic CSI–density pairs from watchable motion

c-resplan-crowd-to-csi · exp-csi-crowd, walk-notebook

Sessions

No sessions yet — this campaign has a brief but no execution.

Brief

Question

Does coupling watchable crowd motion on a real ResPlan floor to the Sionna ray tracer produce synthetic CSI–density pairs that (a) track the crowd and (b) are usable as training data for the BLE-anchored, continuity-regularised decoder? This is the thesis's simulation-to-real backbone — "a JuPedSim→Sionna RT pipeline produces paired synthetic CSI-and-density training data" — a claim with zero supporting papers in the latest synthesis. It builds directly on the motion validated by c-resplan-crowd-geometry-gallery / c-crowd-motion-mix.

What we already know

  • c-csi-crowd-temporal couples jupedsim-runner → sionna-csi-runner, but only on the parametric Geometry-A corridor — jupedsim-runner has no from_floor. The from-floor walker is walk-notebook. So a real-floor coupled chain is walk-notebook (trajectory) → sionna-csi-runner (scene + CSI), a chain that does not exist yet.
  • The 2026-06-15 run-artefact staging (inputs_from, sim/runtime/run_inputs_stage.py) is the mechanism that feeds an upstream run's trajectory.parquet into a downstream run — exactly what a walk-notebook → sionna coupling needs (the same primitive that wires the EXP-S1 reduction).
  • sionna-csi-runner already declares trajectory.parquet as an optional upstream-input and stages its scene from a floor + experiment (Tx/Rx) via scene_stage.stage_scene — so the floor and the crowd share coordinates by construction.
  • c-csi-fidelity-material / c-csi-body-em-fidelity established the ray-traced CSI fidelity envelope; this campaign reuses it on a furnished real layout rather than re-validating the tracer.

Prerequisites (the gating work)

  1. Tx/Rx placement on resplan-12439. stage_scene needs an experiment with experiment_devices (Tx/Rx roles) on the floor. Author one via gis_create_experiment + gis_place_device (a handful of links spanning the living/kitchen/bed rooms), then set exp-csi-crowd.base.experiment to its name. Today the floor has devices for no CSI experiment (experiment_cells == 0).
  2. A walk-notebook → sionna-csi-runner coupled simulation. Either generalise exp-csi-crowd to accept a from_floor walk trajectory as its scatterer source (swap the jupedsim-runner leg for walk-notebook when geometry: from_floor), or define a sibling exp-csi-crowd-floor. The trajectory hand-off uses inputs_from (already built); the scene hand-off uses stage_scene (already built) — the work is the chain wiring, not new staging primitives.
  3. Sionna CPU budget. CPU-only ray tracing on Apple Silicon (no Metal backend) is the cost driver — keep the link count and frame stride modest; run systematic on CI (sim-campaign.yml) rather than the laptop. Hence cpu_hours: 6.0.

What the supervisor does (once prerequisites land)

  1. Phase A — motion. Launch the walk-notebook daily-life runs on resplan-12439 (3 seeds), labelled {campaign: c-resplan-crowd-to-csi, seed}. Reuse c-crowd-motion-mix's validated scenario; confirm the replay + kinematic gate (this is also a free re-validation of the furnished floor).
  2. Phase B — channel. Launch exp-csi-crowd per seed; inputs_from fans in the matching walk-notebook trajectory; stage_scene builds the scene from the floor + Tx/Rx. Produces csi.hdf5 + links.parquet per seed.
  3. Phase C — pairing + reduction. Derive per-cell occupancy from the trajectory, correlate against the CSI link statistics (criterion 2), then run the continuity-penalty reduction on the fanned-in trajectories (criterion 3) to close the loop to the L4½.5 evidence chain.

Figure render request

resplan_crowd_csi — (a) the trajectory.html replay thumbnail beside (b) a CSI-amplitude-vs-time trace per link colour-coded by instantaneous occupancy, (c) the occupancy↔CSI scatter with the fitted monotone relation per band.

Out of scope

  • Hardware-measured CSI. The Sionna channel + the synthetic BLE proxy are simulation; the hardware defence is EXP-F*.
  • Re-validating the ray tracer. Fidelity is c-csi-body-em-fidelity / c-csi-fidelity-material's job; this campaign assumes it.
  • Multi-floor CSI. One furnished floor first; the corpus sweep is a follow-on once the single-floor coupling holds.

Expected interpretation

  1. All criteria met → "Watchable crowd motion on a real floor yields CSI–density pairs that track the crowd and feed the continuity-regularised decoder." First end-to-end instance of the thesis's synthetic-data backbone on real geometry — footnotes the System Design and Supporting Material chapters.
  2. Motion + channel produced, weak occupancy↔CSI correlation → a fidelity/placement finding: the Tx/Rx layout under-illuminates the occupied cells, or the frame stride is too coarse. Re-place devices / refine sampling; cite the scatter.
  3. Coupling does not wire → the prerequisite work (chain definition) is the deliverable; report what the walk-notebook→sionna hand-off needs and defer the science.

Status — 2026-07-06 overhaul triage

Superseded / partially blocked. The chartered purpose — coupled crowd→CSI pairs from watchable motion — has since been delivered by c-csi-crowd-temporal (time-resolved coupled JuPedSim→Sionna), c-feature-type-transfer (coupled pairs across furniture rearrangement, 9 seed-pairs), and c-multifloor-fsm-observatory (full coupled FSM→device-CSI on 6 real floors). The residual novelty here (the python-ml continuity-penalty reduction over the pairs) is blocked on the same missing torch+lightgbm runner as c-exp-s1-mechanism-defense, and the continuity-prior direction was independently demoted (c-continuity-prior-csi-density). Recommend: archive, or re-charter narrowly if a coupled-pairs need arises that the three successor corpora don't cover.