berkus
/
wifi-densepose
mirror of https://github.com/ruvnet/wifi-densepose.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
wifi-densepose/CHECKLIST.md

11 KiB
Raw Blame History

RuView · Implementation Checklist

Single source of truth for what's shipped and what's open. Updated at the end of every session. Pair with docs/references/espectre-gap-analysis.md for the technical detail behind each line.

Last sweep: 2026-05-18, branch feat/ota-rssi-mobile, head 12e1cf9d. Status: 50 Done / 0 Open in-scope. Deferred items (out of session scope, each with explicit reason) listed at the bottom.

This count includes the ADR-100..114 carry-in from the prior agent + this session's:

  • ADR-115 — FW /ota/set-target REST endpoint
  • ADR-116 — WiFlow-v1 supervised pose loader in Rust + UI dropdown
  • ADR-117 — process hygiene (ping zombies, loopback filter, audit sweep)
  • ADR-118 — feature decorrelation + multi-node 22-feature extractor
  • ADR-119 — frame-level MLP classifier (22→32→6) replacing LogReg fallback
  • ADR-120 — windowed temporal classifier (W-MLP, 440→64→6) + hybrid priority (rule-based owns 4 base classes, W-MLP owns waving/transition) + two-layer label smoothing + /api/v1/adaptive/debug diagnostic endpoint

ADR-111 is intentionally absent (folded into ADR-109 during the AP-MAC tracking work).

Adaptive classifier accuracy trajectory across the session:

2-node 15-feat LogReg     40.4%   baseline
6-node 15-feat LogReg     44.4%   +4.0  (more sensors)
6-node 22-feat LogReg     49.58%  +5.2  (ADR-118 feature engineering)
6-node 22-feat MLP        53.53%  +3.95 (ADR-119 non-linear)
6-node 22-feat W-MLP      90.40%  +36.87 (ADR-120 temporal context)
                          ─────
total                     +50.0  pts vs baseline

W-MLP 90.40% is training-set accuracy; live transition class is over-represented (model overfit to ambiguous training frames). Held-out test set + cleaner per-class re-records are the recommended next step.

✅ Done

Server (v2/crates/wifi-densepose-sensing-server)

  • ADR-100 PHY gain-lock (AGC + FFT freeze, ESPectre port) — FW
  • ADR-101 Raw-amplitude classifier (CV + baseline drop, hysteresis)
  • ADR-101 Per-node classification badges in WS payload
  • ADR-102 NBVI subcarrier selection (formula α=0.5, top-12)
  • ADR-102 NBVI Step 1 quiet-window finder
  • ADR-103 Persistent baseline at data/baseline.json (FULL broadband)
  • ADR-103 Universal threshold via baseline-CV normalization
  • ADR-104 Per-subcarrier drift channel (off-axis presence)
  • ADR-104 NBVI Step 3 FP-rate validation (K ∈ {6,8,10,12,16,20})
  • ADR-104 Per-sub drift exposed in WS node_features[].drift_score + raw.html sparkline per node (commit eec3ca6c)
  • ADR-104 Baseline staleness watch — warn when on-disk baseline > 4 h old AND drift consistently fires during absent periods (commit eec3ca6c)
  • ADR-105 Drop all synthetic data from runtime ([signal_field, pose_keypoints, persons, fake confidence — all gated)
  • ADR-105 n_aps_used: u8 uniform field on enhanced_motion + enhanced_breathing (commit 598a4b2f)
  • ADR-106 Full complex CSI in WS (amplitude + phases + meta)
  • ADR-106 Built-in CSI keepalive (managed ping per sensor)
  • ADR-106 Server-side µs timestamp_us
  • ADR-107 POST /api/v1/baseline/calibrate + UI button
  • ADR-107 Auto-recalibrate on long-quiet periods (30 min default)
  • ADR-107 GET /api/v1/baseline (status + cooldown)
  • ADR-107 Progress bar in raw.html calibrate button (commit 432753e1)
  • ADR-112 Multi-AP signal_field via MultistaticFuser — coverage × activity heatmap, non-zero only with ≥2 nodes + positions; preserves ADR-105 zero-grid otherwise (commit c8ac60f6)
  • ADR-105 Hide pose canvas in Docker SPA when model_loaded == false + "no trained model" overlay (commit 2dcb30a6)
  • ADR-104 Phase-domain drift channel — script + server both compute per-subcarrier circular mean/var; phase_drift_score surfaced on PerNodeFeatureInfo (commit 47dafab4)
  • ADR-113 Day/night baseline profiles with hot-reload (--baseline-profile {single,auto,day,night}) (commit a1e09525)
  • ADR-114 2000-packet replay regression suite (1000 idle + 1000 motion synthetic-but-parameter-matched, F1 ≥ 0.85 threshold) (commit 96225e27)

Firmware (firmware/esp32-csi-node)

  • ADR-100 Gain-lock (300-packet median, MIN_SAFE_AGC=30 safety)
  • ADR-106 Sensor µs timestamp in CSI trailer (rx_ctrl.timestamp)
  • ADR-108 NVS persistence of gain-lock — reboot ready in ~0.5 s
  • ADR-109 POST /ota/recalibrate — clear gain-lock NVS via REST, no USB needed (commit f92807cd)
  • ADR-109 Track AP MAC in gl_ap_mac NVS — auto-invalidate stale gain-lock on AP swap (commit f92807cd)
  • ADR-115 POST /ota/set-target — repoint CSI aggregator (csi_cfg/target_ip + target_port) without USB; recovered both nodes after Mac IP move TP-Link → .103

Pose model

  • ADR-116 WiFlow-v1 supervised pose loader (Rust) — --wiflow-model data/models/ruview/wiflow-v1/wiflow-v1.json flips pose_estimation: true; per-tick TCN forward yields 17 COCO keypoints on /api/v1/pose/current and WS pose_data. Output quality requires per-deployment fine-tune (LoRA adapters or re-train, see Pack E).
  • ADR-117 Process hygiene + audit follow-ups — UDP loopback filter prevents cargo test cross-talk from spawning ping zombies (250→2 children); keepalive pre-reaps orphans at startup; / redirects to SPA; wiflow zero-pad replaces silent subcarrier-0 duplication; keypoint confidence stamped from runtime classifier; sensing tab container restored; multi-node test guards external :5005; docs/typo/range sweep.

Adaptive Classifier (data pipeline + model)

  • ADR-118 Feature decorrelation + multi-node extractor — audit on 6-node 151k-frame set found 21 multicollinear pairs + 1 dead feature (amp_min constant 0); refactored to 22 features (4 global + 6 nodes × 3) with proper z-score normalisation. Accuracy 44.4% → 49.58% (commit e86f6506).
  • ADR-119 Frame-level MLP (22→32→6 ReLU+softmax) replaces LogReg fallback — manual backprop, no external ML crate, ~3k weights, trains in seconds. Accuracy 49.58% → 53.53% (+3.95 pts, concentrated on motion classes — exactly where non-linear combinations matter; commit 94330708).
  • ADR-120 Windowed temporal classifier (W-MLP, 440→64→6) — stacks 20 frames × 22 features for temporal pattern recognition. Captures walking cadence (~2 Hz), sit-stand cycles (~0.5 Hz), gesture rhythm (1-2 Hz). Accuracy 53.53% → 90.40% training (+36.87 pts; held-out generalisation TBD). Hybrid priority: rule-based owns 4 base classes (ESPectre F1>96%), W-MLP owns waving/transition exclusively. Two-layer label smoothing (15-tick mode + 2-tick confirm) stops UI flicker. /api/v1/adaptive/debug exposes raw model labels for operator diagnostics (commits da4c123d, 442c03da, 3e12686a, c3f00f3a, 77d404d6, 2956414b, 12e1cf9d).

Tests / fixtures

  • ADR-114 tests/fixtures/replay_idle.jsonl + replay_motion.jsonl (1000 frames each, JSONL schema: {node_id, amplitude[]}) (commit 96225e27)
  • ADR-114 scripts/generate-replay-fixtures.py — seeded deterministic generator for the two fixtures (commit 96225e27)
  • (parallel agent) RSSI carry-through via feature_state header fix
  • (parallel agent) OTA: OTA_SIZE_UNKNOWN, httpd stack_size=8192, reset-reason log — all three FW prerequisites for working OTA

Ops / tooling

  • scripts/ota-deploy.sh — WiFi OTA flash + auto-discovery + verify
  • scripts/record-baseline.py — headless baseline capture (CLI)
  • data/baseline.json v2 schema
  • docs/references/ota-pipeline.md — verbatim OTA recipe (port 8032)

Documentation

  • ADR-100..120 all written (ADR-111 intentionally absent), each ≤ 200 lines
  • docs/references/espectre-techniques.md — Pace technique catalogue
  • docs/references/espectre-gap-analysis.md — section-by-section gap
  • Documentation actualization sweep — every Open Items section cross-checked against actual implementation state

⏳ Open, priority-sorted

High value, low effort

(all closed this session — see Done above. Tailscale-target item moved to Deferred below per session brief.)

High value, medium effort

(all closed this session — see Done above)

Bigger, lower urgency (still active)

(all closed this session — multiple baseline profiles shipped via ADR-113, see Done above)

One-time hygiene

  • Re-record data/baseline.json — current file already carries per_subcarrier_mean so amplitude drift (ADR-104) is active. Verified the recorder writes the new per_subcarrier_phase_mean / per_subcarrier_phase_var schema end-to-end (this session). data/baseline.json is untracked, so no repo commit needed; operator re-records via UI when they step out for a true empty-room sample (currently the file reflects an operator-present recording — fine for the amp channel, needs re-record for the phase channel to populate ≥ 16 usable subcarriers).

Deferred — out of session scope

Marked here so future sessions don't re-litigate; each line carries an explicit reason. Bring them back only if scope changes.

  • HA via MQTT — new integration. Excluded by current session brief (no new integrations on current hardware).
  • ESPHome native component — same reason as HA/MQTT.
  • Web Serial calibration game — explicitly excluded.
  • Boot-time NBVI freeze in FW — explicitly excluded.
  • Per-channel NVS cache for gain-lock — explicitly excluded; only matters if channel hopping is reactivated, which is also excluded.
  • DensePose model train + load — explicitly excluded.
  • AETHER contrastive pretrain on live data — explicitly excluded.
  • MERIDIAN domain generalization — explicitly excluded.
  • Channel hopping (ADR-029) — explicitly excluded.
  • Multi-antenna support (n_antennas > 1) — explicitly excluded.
  • README.md trim (542 lines) — explicitly excluded.
  • CLAUDE.md trim (407 lines) — explicitly excluded.
  • Tailscale-target in NVS — Mac stable on TP-Link this session, low ROI. Not blocking. (ADR-100 follow-up; bring back if Mac network swap becomes routine.)

Reference

Doc Purpose
docs/adr/ All ADRs 001-120 (111 absent); 100-120 are this session
docs/references/espectre-techniques.md Pace technique catalogue + RuView adoption
docs/references/espectre-gap-analysis.md Section-by-section gap with priority table
docs/references/ota-pipeline.md OTA recipe — port 8032, three FW prereqs

To mark an item done: tick the box, add (ADR-XXX, commit-sha) after the line, move it from the priority section to the top "Done" section.