v0.6.2-esp32 — ADR-081 Adaptive CSI Mesh Kernel + Timer Svc stack fix

First firmware release of the ADR-081 Adaptive CSI Mesh Kernel, plus the Timer Service stack-overflow regression fix caught by on-device validation. Ships both 8 MB and 4 MB variants from CI.

Highlights

• 5-layer adaptive CSI kernel (Radio Abstraction → Adaptive Controller → Mesh Plane → Feature State → Rust handoff) — swap radio silicon without touching the Rust signal/ruvector/mat crates
• 60-byte rv_feature_state_t packet (magic 0xC5110006) replaces raw CSI as the default upstream — ~99.7 % bandwidth reduction (300 B/s at 5 Hz vs ~100 KB/s raw)
• Mesh sensing plane (rv_mesh) with 4 roles, 7 message types, CRC32-integrity envelope — HEALTH every 30 s, ANOMALY_ALERT on state transitions
• Host-testable pure decision policy (adaptive_controller_decide.c) — 33/33 assertions, decide() at 3.2 ns/call
• 4 MB variant (ESP32-S3 SuperMini) now built in CI alongside the 8 MB default

Validated on real hardware

ESP32-S3 QFN56 rev v0.2 (MAC 3c:0f:02:e9:b5:f8), 38 s continuous run after flash:

• App version: 0.6.2 ✅
• 0 stack-overflow / panic / reboot markers
• 28 adaptive_ctrl medium ticks, steady state, no drift
• 149 rv_feature_state_t emissions at ~5 Hz on earlier 30 s validation
• Real WiFi CSI: 37 – 72 pps yield, RSSI -43 to -57 dBm

Fixes

Fix	Detail
Timer Svc stack overflow (new)	emit_feature_state() + stream_sender network I/O runs inside FreeRTOS Timer Svc; ESP-IDF default 2 KiB stack overflows ~1 s after boot. Bumped CONFIG_FREERTOS_TIMER_TASK_STACK_DEPTH=8192 in all three sdkconfig.defaults*. Follow-up: move heavy work out of the timer daemon.
adaptive_controller.c implicit decl (#404)	fast_loop_cb called emit_feature_state() before its static definition, tripping -Werror=implicit-function-declaration. Added forward declaration.
provision.py esptool v5 compat (#391)	write-flash (hyphenated) in dry-run hint for esptool ≥ 5.
provision.py silent NVS wipe (#391)	Now refuses to run without --ssid, --password, --target-ip unless --force-partial is passed; --force-partial prints which keys will be wiped.
Defensive node_id capture (#232, #375, #385, #386, #390)	csi_collector_init() now captures g_nvs_config.node_id once into a module-local; a canary WARN fires on divergence.

CI change

firmware-ci.yml is now a matrix job building both variants in parallel and uploading variant-named artifacts — 6-binary releases no longer need local ESP-IDF.

Installing

Six binaries are attached below. The default flash layout for ESP32-S3 8 MB:

esptool.py --chip esp32s3 --port /dev/ttyUSB0 --baud 460800 \
  --before default_reset --after hard_reset \
  write_flash -z --flash_mode dio --flash_freq 80m --flash_size 8MB \
  0x0      bootloader.bin \
  0x8000   partition-table.bin \
  0xf000   ota_data_initial.bin \
  0x20000  esp32-csi-node.bin

For ESP32-S3 SuperMini 4 MB, use esp32-csi-node-4mb.bin at 0x20000 and partition-table-4mb.bin at 0x8000 (reuse the same bootloader and OTA data).

After flash, provision WiFi + target (replaces the full csi_cfg NVS namespace — pass all creds at once):

python firmware/esp32-csi-node/provision.py \
  --port /dev/ttyUSB0 \
  --ssid <SSID> --password <PSK> --target-ip <DEST_IP>

Artifacts

File	Size	Target
esp32-csi-node.bin	1.06 MB	8 MB app (flash @ 0x20000)
esp32-csi-node-4mb.bin	851 KB	4 MB app (flash @ 0x20000)
bootloader.bin	18 KB	Both (flash @ 0x0)
partition-table.bin	3 KB	8 MB layout (flash @ 0x8000)
partition-table-4mb.bin	3 KB	4 MB layout (flash @ 0x8000)
ota_data_initial.bin	8 KB	OTA data (flash @ 0xf000)

Compatibility

• ESP-IDF: v5.4
• Target: ESP32-S3 (Xtensa dual-core). Not supported: ESP32 (original), ESP32-C3 (single-core, can't run CSI DSP).
• Supersedes: v0.6.1-esp32
• Wire format: rv_feature_state_t magic 0xC5110006 is the new default upstream; raw ADR-018 CSI (magic 0xC5110001) remains available as a debug stream gated by the channel plan.

Reference

• ADR-081 — Adaptive CSI Mesh Firmware Kernel (docs/adr/ADR-081-adaptive-csi-mesh-firmware-kernel.md)
• Full changelog entry: CHANGELOG.md#v062-esp32
• Source tag: ae40e2b

ESP32-S3 CSI Firmware v0.6.1 — Defensive node_id Fix

ESP32-S3 CSI Firmware v0.6.1

Critical fix: node_id clobber on multi-node deployments

Closes #232, #375, #385, #386, #390.

Users on multi-node deployments reported that every ESP32 node transmitted node_id=1 in UDP frames, despite NVS provisioning showing the correct value. This made multi-node setups indistinguishable downstream.

What changed

Defense-in-depth fix — the firmware now captures node_id into a module-local static at init time, isolating the UDP frame header from any memory corruption of g_nvs_config:

1. csi_collector_init() captures g_nvs_config.node_id into static uint8_t s_node_id once at init
2. csi_serialize_frame() writes buf[4] = s_node_id — bypasses the global entirely
3. All other consumers (edge_processing.c, wasm_runtime.c, display_ui.c, swarm_bridge_init) use the new csi_collector_get_node_id() accessor
4. Clobber canary logs WARN if g_nvs_config.node_id diverges from the captured value at init — helps isolate the upstream corruption path

Hardware validation (ESP32-S3 on COM8)

Check	Result
NVS node_id	2 (provisioned)
Boot log main:	Node ID: 2
Defensive capture log	Captured node_id=2 at init
csi_collector init log	node_id=2, channel=5
UDP packets (15/15)	byte[4] = 2
Clobber canary	Quiet (no WARN)
App version	0.6.1
Binary size	865 KB (54% flash free)

Binaries

File	Flash size	Description
esp32-csi-node.bin	8MB	Main firmware
bootloader.bin	8MB	Bootloader
partition-table.bin	8MB	Partition table (dual OTA)
ota_data_initial.bin	8MB	OTA data
esp32-csi-node-4mb.bin	4MB	Main firmware (SuperMini)
partition-table-4mb.bin	4MB	Partition table (SuperMini)

Flash command

python -m esptool --chip esp32s3 -p COM8 -b 460800 \
  --before default_reset --after hard_reset \
  write_flash --flash_mode dio --flash_size 8MB --flash_freq 80m \
  0x0 bootloader.bin \
  0x8000 partition-table.bin \
  0xf000 ota_data_initial.bin \
  0x20000 esp32-csi-node.bin

For 4MB SuperMini boards, use esp32-csi-node-4mb.bin and partition-table-4mb.bin with --flash_size 4MB.

SHA-256 hashes

7f943b26...88197f7  bootloader.bin
9e650bf1...54bd43   esp32-csi-node.bin
15c263f6...02b32    esp32-csi-node-4mb.bin
7d2c7ac4...82c62f   ota_data_initial.bin
4c2cc4ff...d81f0    partition-table.bin
4c2cc4ff...d81f0    partition-table-4mb.bin

Who should upgrade

All users running multi-node ESP32-S3 deployments. Single-node users are unaffected but the upgrade is safe.

v0.7.0 — WiFlow Camera-Supervised Pose Model (92.9% PCK@20)

WiFlow v1 — Camera-Supervised WiFi Pose Estimation

First WiFlow model trained on real ESP32 CSI + real camera ground truth.

Headline Result

Metric	Value
PCK@20	92.9%
Eval loss	0.082
Bone constraint	0.008
Parameters	186,946
Model size	974 KB
Training time	19 minutes

What's New

Camera Ground-Truth Pipeline (ADR-079)

• collect-ground-truth.py — MediaPipe webcam capture synced with CSI
• align-ground-truth.js — Nanosecond time alignment
• train-wiflow-supervised.js — 3-phase curriculum training
• eval-wiflow.js — PCK/MPJPE evaluation
• record-csi-udp.py — Lightweight ESP32 CSI recorder

ruvector Optimizations

• O6: Subcarrier selection (70→35, 50% reduction)
• O7: Attention-weighted subcarriers
• O8: Stoer-Wagner min-cut person separation
• O9: Multi-SPSA gradient estimation
• O10: Scalable model (lite/small/medium/full)

Data Collection

• ESP32-S3 CSI: 7,000 frames at 23fps (5 min)
• Mac Mini M4 Pro camera: 6,470 frames via MediaPipe (5 min)
• 345 time-aligned paired samples

Model Archive Contents

• wiflow-v1.json — Trained model weights (974 KB)
• training-log.json — Loss curves per phase
• baseline-report.json — Pre-training baseline metrics
• MODEL_CARD.md — Model documentation

How to Use

# Collect your own ground truth
python scripts/collect-ground-truth.py --duration 300 --preview
python scripts/record-csi-udp.py --duration 300

# Train
node scripts/train-wiflow-supervised.js --data data/paired/your-data.jsonl --scale lite

# Evaluate
node scripts/eval-wiflow.js --model models/wiflow-real/wiflow-v1.json --data data/paired/your-data.jsonl

Related

• PR #363: Camera ground-truth training pipeline
• ADR-079: Camera Ground-Truth Training Pipeline
• ADR-072: WiFlow Architecture

v0.6.0 — Pre-Trained Models on HuggingFace + 17 Sensing Apps

Pre-trained weights: https://huggingface.co/ruv/ruview | mirror

0.008ms inference | 164K emb/s | 100% presence | 51.6% contrastive improvement | 8 KB model

Trained on 60,630 overnight samples. 17 sensing applications. 10 ADRs (069-078).

New: sleep monitor, apnea detector, stress monitor, gait analyzer, RF tomography, passive radar, material classifier, through-wall detector, device fingerprint.

PR #341 (Kalman tracker by @taylorjdawson) + PR #310 (security fix) merged.

HuggingFace | Cognitum.one | Tutorial

ESP32-S3 CSI Firmware v0.5.5 — Advanced Sensing + WiFlow + Camera-Free Training

v0.5.5 — Advanced Sensing, WiFlow Architecture, Camera-Free Pose Training

TL;DR

Your $27 sensor kit now has spiking neural networks that learn your room in 30 seconds, min-cut person counting that actually works (fixes #348), CNN spectrogram embeddings for environment fingerprinting, multi-frequency mesh scanning across 6 WiFi channels, and the WiFlow SOTA architecture (1.8M params) for 17-keypoint pose estimation — all trainable without a camera.

What Changed (v0.5.4 → v0.5.5)

Feature	v0.5.4	v0.5.5
Person counting	Broken (always 4)	MinCut: correct count (fixes #348)
WiFi channels	1 (ch 5 only)	6 channels (hopping 1/3/5/6/9/11)
Pose model	Simple 8→64→128 encoder	WiFlow: TCN + axial attention (1.8M params)
Pose keypoints	5 proxy → 17 interpolated	17 COCO keypoints (WiFlow decoder)
Online learning	None	SNN with STDP (adapts in <30s)
Environment fingerprint	8-dim features	128-dim CNN spectrogram
Multi-node fusion	Average	Graph transformer (GATv2 attention)
RF scanning	None	Live spectrum visualization
Neighbor WiFi	Ignored	Used as passive radar illuminators
Training pipeline	ruvllm basic	+ camera-free + WiFlow + GCloud + Mac Mini
ADRs	069-071	069-076 (8 total)

New Capabilities

ADR-073: Multi-Frequency Mesh Scanning

ESP32 nodes hop across channels 1/3/5/6/9/11 at 200ms dwell. Neighbor WiFi networks (your printer, router, neighbors) become free RF illuminators. Null subcarriers dropped from 19% to 16%.

# Provision channel hopping
python firmware/esp32-csi-node/provision.py --port COM9 \
  --hop-channels "1,6,11" --hop-dwell 200

# Live RF scan
node scripts/rf-scan.js --port 5006 --duration 30

ADR-074: Spiking Neural Network

128→64→8 SNN with STDP unsupervised learning. Adapts to room in <30s without labels. 16-160x less compute than the FC encoder (event-driven, only processes changes).

node scripts/snn-csi-processor.js --port 5006

ADR-075: MinCut Person Separation (fixes #348)

Stoer-Wagner min-cut on subcarrier correlation graph. Correctly counts 1 person on all 24 test windows where old firmware showed 4. <5ms per window.

node scripts/mincut-person-counter.js --port 5006
# Or replay: node scripts/mincut-person-counter.js --replay data/recordings/*.csi.jsonl

ADR-076: CNN Spectrogram Embeddings + Graph Transformer

CSI 64×20 matrix → 224×224 grayscale → CNN → 128-dim embedding. Same-node similarity 0.95+. GATv2 multi-head attention fuses multi-node features.

node scripts/csi-spectrogram.js --replay data/recordings/*.csi.jsonl
node scripts/mesh-graph-transformer.js --port 5006

ADR-072: WiFlow SOTA Architecture

Full reimplementation of WiFlow (arXiv:2602.08661) in pure JS:

• TCN temporal encoder (dilated causal conv, k=7)
• Asymmetric spatial encoder (1×3 residual blocks)
• Axial self-attention (8 heads, width + height)
• Pose decoder → 17 COCO keypoints
• SmoothL1 + bone constraint loss (14 skeleton connections)
• 1.8M parameters (881 KB at 4-bit quantization)

Camera-Free Training (ADR-071 extended)

10 sensor signals replace cameras: PIR, BME280 temp/humidity, RSSI triangulation, subcarrier asymmetry, vibration, reed switch, kNN clusters, boundary fragility.

5-phase pipeline: multi-modal collection → weak labels → 5-keypoint proxy → 17-keypoint interpolation → self-refinement.

Validated Benchmarks

Metric	Value
Rust tests	1,463 passed
Presence accuracy	100%
MinCut person count	24/24 correct (was 0/24)
Inference latency	0.012 ms (M4 Pro)
Throughput	171,472 emb/s
WiFlow PCK@20	2.5% (camera-free baseline)
WiFlow parameters	1,804,962
WiFlow model (4-bit)	881 KB
CNN spectrogram similarity	0.95+ (same-node)
SNN adaptation time	<30s (STDP online)
Channel hopping	6 channels, 200ms dwell
Null reduction	19% → 16%

Flash Instructions

Same firmware binary as v0.5.4 (channel hopping is NVS-configured, no reflash needed for existing v0.5.4 users):

python -m esptool --chip esp32s3 --port COM9 --baud 460800 \
  write_flash --flash-mode dio --flash-size 8MB --flash-freq 80m \
  0x0 bootloader.bin 0x8000 partition-table.bin \
  0xf000 ota_data_initial.bin 0x20000 esp32-csi-node.bin

# Enable channel hopping (new in v0.5.5)
python firmware/esp32-csi-node/provision.py --port COM9 \
  --hop-channels "1,6,11" --hop-dwell 200

New Scripts (15+)

Script	Purpose
rf-scan.js	Live RF spectrum scanner
rf-scan-multifreq.js	Multi-channel wideband view
snn-csi-processor.js	Spiking neural network processor
mincut-person-counter.js	Correct person counting
csi-graph-visualizer.js	Correlation graph visualization
csi-spectrogram.js	CNN spectrogram embeddings
mesh-graph-transformer.js	GATv2 multi-node fusion
train-wiflow.js	WiFlow SOTA training
train-camera-free.js	Camera-free 17-keypoint training
train-ruvllm.js	ruvllm contrastive + LoRA + TurboQuant
benchmark-ruvllm.js	Model benchmarking
benchmark-wiflow.js	WiFlow benchmarking
benchmark-rf-scan.js	RF scan benchmarking
wiflow-model.js	WiFlow architecture (pure JS)
seed_csi_bridge.py	ESP32 → Cognitum Seed ingest

Learn more: Cognitum.one · Tutorial · User Guide

Full Changelog: v0.5.4-esp32...v0.5.5-esp32

---

Session Statistics

Metric	Value
Commits	24
Files changed	74
Lines added	22,231
ADRs	8 (069-076)
Research docs	26
New scripts	20+
PRs merged	2 (#350, #352)
Issues addressed	4 (#348, #188, #190, #268)
Rust tests	1,463 passed
WiFlow PCK@20	2.5% (camera-free, improving with more data)
Overnight data	65,938+ frames collecting
Session cost	~$0.50 (Mac Mini electricity)

ESP32-S3 CSI Firmware v0.5.4 — Cognitum Seed Integration

ESP32-S3 CSI Firmware v0.5.4 — Cognitum Seed Integration

TL;DR

Your ESP32 can now remember what it senses. By connecting to a Cognitum Seed every presence detection, breathing measurement, and motion event gets stored as a searchable vector with cryptographic proof it happened.

What Changed (v0.5.3 → v0.5.4)

Feature	v0.5.3	v0.5.4
CSI sensing	✅ Real-time only	✅ Real-time + persistent storage
Data retention	❌ Lost on reboot	✅ RVF vector store on Cognitum Seed
Similarity search	❌ None	✅ kNN queries ("find states like this one")
Data integrity	❌ None	✅ SHA-256 witness chain + Ed25519 signing
AI integration	❌ None	✅ 114-tool MCP proxy (Claude, GPT can query sensors)
Environmental sensors	❌ CSI only	✅ + BME280, PIR, vibration, ADC (via Seed)
Feature vectors	❌ Raw only	✅ 8-dim normalized @ 1 Hz (new packet 0xC5110003)
Security	Basic	✅ Bearer tokens, TLS, source filtering, NaN rejection

New: 8-Dimension Feature Vector

Every second, the ESP32 sends a compact 48-byte packet summarizing what it sees:

Dim	What it measures	Example
0	Is someone there? (presence)	0.45 = maybe, 0.95 = definitely
1	Are they moving? (motion)	0.0 = still, 0.87 = active
2	Breathing rate	0.69 = 20.8 BPM (normal)
3	Heart rate	0.75 = 90 BPM (normal)
4	Signal quality (phase variance)	Higher = more activity
5	How many people?	0.25 = 1 person, 0.50 = 2
6	Fall detected?	0.0 = no, 1.0 = yes
7	Signal strength (RSSI)	0.54 = -46 dBm (good)

These vectors are stored on the Cognitum Seed and can be searched: "Find the 10 most similar states to right now" — useful for anomaly detection, pattern recognition, and environment fingerprinting.

Architecture

ESP32-S3 ($9)  ──UDP──>  Your laptop  ──HTTPS──>  Cognitum Seed ($15)
  WiFi CSI capture          Bridge script            Stores vectors forever
  28 Hz raw frames          Batches & validates      kNN similarity search
  1 Hz feature vectors      NaN rejection            Witness chain (proof)
  1 Hz vital signs          Source filtering          114-tool AI proxy

How to Use

# 1. Flash firmware to ESP32-S3
python -m esptool --chip esp32s3 --port COM9 --baud 460800 \
  write_flash --flash-mode dio --flash-size 8MB --flash-freq 80m \
  0x0 bootloader.bin 0x8000 partition-table.bin \
  0xf000 ota_data_initial.bin 0x20000 esp32-csi-node.bin

# 2. Set WiFi + target (your laptop IP)
python firmware/esp32-csi-node/provision.py --port COM9 \
  --ssid "YourWiFi" --password "secret" \
  --target-ip 192.168.1.20 --target-port 5006

# 3. Pair with Cognitum Seed (plug in via USB first)
curl -sk -X POST https://169.254.42.1:8443/api/v1/pair/window
curl -sk -X POST https://169.254.42.1:8443/api/v1/pair \
  -H 'Content-Type: application/json' -d '{"client_name":"my-laptop"}'

# 4. Run the bridge
export SEED_TOKEN="<token-from-step-3>"
python scripts/seed_csi_bridge.py \
  --seed-url https://169.254.42.1:8443 --token "$SEED_TOKEN" --validate

# 5. Check what's stored
python scripts/seed_csi_bridge.py --token "$SEED_TOKEN" --stats

No Cognitum Seed? The ESP32 firmware works exactly as before — all existing features (CSI streaming, vitals, presence, fall detection, mesh TDM) are unchanged. The Seed integration is additive.

Learn more: Cognitum.one · ADR-069 · User Guide

Validation Results

Test	Result
Rust workspace	1,463 passed, 0 failed
Python proof	VERDICT: PASS
Firmware size	844 KB (55% flash free)
Feature vectors	1 Hz, 8-dim, all in [0.0, 1.0]
Vitals ↔ Features	HR delta=0.0, BR delta=0.0
Seed ingest	941 vectors, 100% kNN match
Witness chain	SHA-256 verified (1,325 entries)
Regression	6/6 PASSED — NO REGRESSIONS

Binary Hashes (SHA-256)

cf05de2b  esp32-csi-node.bin   (844 KB)
cbb8e3ba  bootloader.bin       (19 KB)
4c2cc4ff  partition-table.bin  (3 KB)
7d2c7ac4  ota_data_initial.bin (8 KB)

Full Changelog: v0.5.3-esp32...v0.5.4-esp32

ESP32-S3 CSI Firmware v0.5.3 — Cross-Node Fusion + DynamicMinCut

What's Changed

Cross-Node Mesh Fusion

• RSSI-weighted feature fusion across multiple ESP32 nodes — closer node gets higher weight
• DynamicMinCut person separation via ruvector_mincut on subcarrier temporal correlation graph
• RSSI-based skeleton position tracking — walk between nodes and the skeleton follows
• Motion-responsive skeleton — arm swing (0-80px), leg kick (0-50px), walking bob, head nod all driven by real CSI data

Per-Node State (ADR-068)

• Each ESP32 gets independent classification, vitals, person count
• Stale node eviction after 60s inactivity
• Multi-node aggregation: max(per_node_counts)

RuVector Signal Processing (Phases 1-3)

• Subcarrier importance weighting via mincut partition
• Temporal EMA keypoint smoothing with coherence gating
• Skeleton kinematic constraints (Jakobsen relaxation, 12-bone COCO-17 tree)
• Compressed pose history for trajectory matching

Smoothing & Quality

• Client-side lerp interpolation (alpha=0.15) for fluid skeleton
• Server EMA (alpha=0.15) for temporal blending
• Dampened procedural animation, recalibrated for real ESP32 data
• Vital sign filtering: larger median window (31), faster EMA, looser jump filter

Hardware Benchmarks (2 ESP32 nodes, 30s)

Metric	v0.5.2 (baseline)	v0.5.3 (fusion)	Improvement
Keypoint jitter	std=4.5px	std=1.3px	-72%
Feature noise	109.4	77.6	-29%
Feature stability	std=154.1	std=105.4	-32%
Confidence	0.643	0.686	+7%
Presence accuracy	93.4%	94.6%	+1.3pp

Bug Fixes

• RSSI byte offset (#332), stack overflow, stale node leak, unsafe pointer removed
• Person count: sum→max aggregation, recalibrated thresholds
• Firmware CI: IDF v5.4, xxd→od

Binaries

File	Flash Size	Description
esp32-csi-node.bin	8MB	Main firmware
esp32-csi-node-4mb.bin	4MB	For 4MB flash boards
bootloader.bin	—	Bootloader
partition-table.bin	8MB	Partition table
partition-table-4mb.bin	4MB	Partition table (4MB)
ota_data_initial.bin	—	OTA data

Full Changelog: v0.5.2-esp32...v0.5.3-esp32

ESP32-S3 CSI Firmware v0.5.2 — RSSI Fix + Per-Node State

What's Changed

Bug Fixes

RSSI byte offset mismatch (#332)

• Server was reading RSSI from byte 14 (sequence counter high byte) instead of byte 16 (actual RSSI)
• Root cause: firmware packs n_subcarriers as u16 and freq_mhz as u32, but server assumed u8 and u16
• Result: RSSI now shows correct negative dBm values (e.g., -46) instead of bogus positive values (+9)

Per-node state pipeline (#249, ADR-068)

• Each ESP32 node now gets independent classification, person count, and vital sign tracking
• Fixes the #1 user-reported issue: detection showing same output regardless of number of nodes/people
• Scales to 256 nodes with <13 MB memory overhead

Firmware CI fixed (#327)

• Upgraded from ESP-IDF v5.2 → v5.4, replaced xxd with od
• Pre-compiled binaries now built and uploaded automatically

Sensing Server

Rebuild from source to get all fixes:

git pull origin main
cd rust-port/wifi-densepose-rs
cargo build --release -p wifi-densepose-sensing-server

Firmware

No firmware binary changes from v0.5.1 — the RSSI fix and per-node state are server-side only. If you already flashed v0.5.1, no reflash needed.

Full Changelog: v0.5.1-esp32...v0.5.2-esp32

ESP32-S3 CSI Firmware v0.5.1 — Watchdog Fix + Edge Detection

What's Changed

Bug Fixes

Firmware: Watchdog crash on busy LANs (#321)

• edge_dsp task now batch-limits to 4 frames before a 20ms yield
• Fixed idle-path busy-spin where pdMS_TO_TICKS(5)==0 at 100Hz tick rate
• Added ring buffer drop counter (drops=N in serial log) for diagnostics
• QEMU validated: All 11 CI jobs pass (7 configs + fuzz + swarm + NVS + build)

Sensing Server: No detection from edge vitals (#323)

• ESP32 nodes running Tier 2+ edge DSP now trigger full sensing_update in the UI
• Previously, vitals packets (magic 0xC5110002) were broadcast but never generated person detection
• Breathing animation now works from edge vitals (non-zero breathing_band_power)

Python Package: ModuleNotFoundError (#314)

• Added wifi_densepose/ package with WiFiDensePose facade class
• from wifi_densepose import WiFiDensePose now works as documented in README

CI Validation

Job	Status
Build Espressif QEMU	✅
NVS Matrix Generation	✅
Fuzz Testing (ADR-061 Layer 6)	✅
Swarm Test (ADR-062)	✅
QEMU Test — default	✅
QEMU Test — edge-tier0	✅
QEMU Test — edge-tier1	✅
QEMU Test — tdm-3node	✅
QEMU Test — boundary-min	✅
QEMU Test — boundary-max	✅
QEMU Test — full-adr060	✅

Upgrade Notes

• Firmware: Reflash required for the watchdog fix. Build from source using ESP-IDF v5.4.
• Sensing server: Rebuild from main — cargo build --release -p wifi-densepose-sensing-server
• Python: pip install -e . from repo root, or from wifi_densepose import WiFiDensePose

Full Changelog: v0.5.0-esp32...v0.5.1-esp32

ESP32-S3 CSI Firmware v0.5.0 — mmWave Sensor Fusion

ESP32-S3 CSI Firmware v0.5.0 — mmWave Sensor Fusion (ADR-063/064)

What's New

60 GHz mmWave Radar Fusion — The firmware now auto-detects mmWave radar modules connected via UART and fuses their data with WiFi CSI for dramatically improved sensing.

Supported mmWave sensors:

Sensor	Frequency	Capabilities	Cost
Seeed MR60BHA2	60 GHz	Heart rate, breathing, presence, distance	~$15
HLK-LD2410	24 GHz	Presence, distance (motion + static)	~$3

Auto-detection: The firmware probes UART1 (GPIO17/18) at boot — first at 115200 baud (MR60BHA2), then 256000 baud (LD2410). If a sensor is found, it registers capabilities and starts a background parsing task. No configuration needed.

48-byte fused vitals packet (magic 0xC5110004): When mmWave is active, vitals are fused using weighted Kalman averaging (mmWave 80% + CSI 20%). Falls back to the standard 32-byte CSI-only packet when no mmWave is detected — fully backward compatible.

Server-side fusion bridge (scripts/mmwave_fusion_bridge.py): For setups where the mmWave runs on a separate ESP32 (e.g., ESP32-C6 with ESPHome), this script reads both serial ports and fuses data in real-time.

Hardware Verified

Dual-sensor live capture (30 seconds, 2026-03-15):

• COM7 (ESP32-S3): WiFi CSI on channel 5, RSSI -41 dBm
• COM4 (ESP32-C6 + MR60BHA2): HR 75 bpm, BR 25/min, person at 52 cm
• Both sensors feeding data concurrently

Also Includes (from v0.4.3.1)

• Fall detection fix: threshold 15.0 rad/s², 3-frame debounce, 5s cooldown (#263)
• Task watchdog fix: vTaskDelay after every frame (#266)
• 4MB flash support (#265)
• All 11 QEMU CI jobs green

Download Guide

File	Flash	Size
esp32-csi-node.bin	8MB	990 KB
bootloader.bin	Both	18 KB
partition-table.bin	8MB	3 KB
ota_data_initial.bin	Both	8 KB
esp32-csi-node-4mb.bin	4MB	773 KB
partition-table-4mb.bin	4MB	3 KB

Flash — 8MB:

python -m esptool --chip esp32s3 --port COM7 --baud 460800 \
  write_flash --flash_mode dio --flash_size 8MB --flash_freq 80m \
  0x0 bootloader.bin 0x8000 partition-table.bin \
  0xf000 ota_data_initial.bin 0x20000 esp32-csi-node.bin

Flash — 4MB:

python -m esptool --chip esp32s3 --port COM7 --baud 460800 \
  write_flash --flash_mode dio --flash_size 4MB --flash_freq 80m \
  0x0 bootloader.bin 0x8000 partition-table-4mb.bin \
  0xF000 ota_data_initial.bin 0x20000 esp32-csi-node-4mb.bin

Provision:

python firmware/esp32-csi-node/provision.py --port COM7 \
  --ssid "YourWiFi" --password "YourPassword" --target-ip 192.168.1.20

Full Changelog

See CHANGELOG.md

Closes #269