ADR 0049: CSI Phase-Based Motion Detection Exploration

Status: Proposed (Exploratory) Date: 2025-12-23 Authors: Steve (mlehaptics), Claude Code (Opus 4.5) Related: AD048 (ESP-NOW Transport), RFIP Addendum (ranging)

Context

Our time synchronization architecture successfully corrects for path asymmetry using statistical methods: EMA filtering, Kalman prediction, outlier rejection, and drift rate estimation. The same mathematical framework could potentially apply to phase-based RF measurements from WiFi Channel State Information (CSI).

802.11mc FTM provides meter-level ranging via Time-of-Flight. Phase-based interferometry could theoretically achieve sub-centimeter precision—the physics is sound (2.4 GHz → 12.5cm wavelength). The challenge is that raw phase measurements are corrupted by:

Corruption Source	Signature	Analogous Timing Problem
Carrier Frequency Offset (CFO)	Linear phase rotation vs time	Clock drift rate
Sampling Time Offset (STO)	Linear phase slope vs subcarrier	RTT asymmetry
PLL jitter	High-frequency phase noise	BLE jitter
Packet sync reset	Random phase jump per packet	Connection event timing

These corruptions have predictable signatures that can be modeled and subtracted—exactly as we model and subtract timing asymmetry.

Decision

Establish a phased exploration to evaluate whether CSI phase sanitization is practical on ESP32-C6 hardware, with the ultimate goal of enhancing RFIP ranging precision or enabling new capabilities (relative motion detection, proximity sensing).

Phase 1: Data Collection (Current Priority)

Goal: Capture raw CSI data and build intuition.

Implementation: 1. Enable CSI collection in ESP-IDF via esp_wifi_set_csi_rx_cb() 2. Log amplitude + phase for all subcarriers during ESP-NOW exchanges 3. Visualize data while moving devices (serial plotter or CSV export) 4. Document observed patterns: noise floor, CFO rotation rate, STO slope

Deliverable: Raw CSI logging integrated into ESP-NOW transport layer with optional enable flag.

No runtime processing in Phase 1 - just data collection for offline analysis.

Phase 2: Offline Analysis

Goal: Characterize corruption signatures and evaluate correctability.

• Measure CFO rate (phase rotation vs time) and stability
• Measure STO slope (phase vs subcarrier) and consistency
• Quantify noise floor after subtracting CFO/STO
• Determine if residual phase correlates with device distance/motion

Deliverable: Analysis report with go/no-go recommendation for Phase 3.

Phase 3: Runtime Prediction Model

Goal: Implement real-time phase sanitization analogous to timing asymmetry correction.

• CFO tracker: Kalman filter on phase rotation rate
• STO estimator: Linear regression across subcarriers
• Motion detector: Track sanitized phase derivative
• Optional: Fuse with FTM ToF for integer ambiguity resolution

Deliverable: csi_phase_tracker_t module with runtime correction.

Phase 4: Integration

Goal: Determine practical applications.

Possible outcomes: - Enhanced RFIP: Sub-centimeter relative positioning between paired devices - Motion detection: Detect approaching/departing motion for proximity features - Gesture sensing: Evaluate feasibility for therapy device interaction - Negative result: Document why phase-based methods aren't practical for this hardware/environment

Rationale

Why Explore This

1. Mathematical parallel: Our asymmetry correction proves we can build runtime statistical models for systematic RF errors. Phase corruption is the same class of problem.

2. Hardware capability: ESP32 exposes CSI data. The capability exists; we're just not using it.

3. Potential payoff: Sub-centimeter ranging would enable applications impossible with meter-level FTM (precise formation control, collision avoidance, docking).

4. Low-risk exploration: Phase 1 is just data logging. If the data looks unpromising, we stop.

Why Phase 1 is Data-Only

• Current development is in a critical work phase (P7.1 pattern playback)
• Exploration should not destabilize production code
• Offline analysis avoids runtime overhead during evaluation
• "Measure twice, cut once" - understand the data before writing algorithms

Technical Notes

ESP-IDF CSI API

wifi_csi_config_t csi_config = {
    .lltf_en = true,           // Long Training Field - most stable for phase
    .htltf_en = true,          // HT Long Training Field
    .stbc_htltf2_en = true,    // STBC HT-LTF
    .ltf_merge_en = true,      // Merge multiple LTF
    .channel_filter_en = false,// Raw data preferred
    .manu_scale = false,       // Automatic scaling
};
esp_wifi_set_csi_config(&csi_config);
esp_wifi_set_csi_rx_cb(csi_rx_callback, NULL);

CSI Data Format

Each callback provides: - RSSI and noise floor - ~52 complex values (amplitude + phase) for OFDM subcarriers - Timestamp - Source MAC (identifies which peer)

Phase Representation

CSI data arrives as int8_t pairs (I/Q). Phase = atan2(Q, I). Phase wraps at ±π.

References

1. ESP-IDF Programming Guide: Wi-Fi CSI
2. Halperin et al., "Tool Release: Gathering 802.11n Traces with Channel State Information," ACM CCR 2011
3. Wang et al., "Understanding and Modeling of WiFi Signal Based Human Activity Recognition," MobiCom 2015
4. GitHub: ESP32-CSI-Tool (StevenMHernandez)
5. AD043: Filtered Time Synchronization (asymmetry correction prior art)

Success Criteria

• Phase 1 complete: CSI data logged, visualized, patterns documented
• Phase 2 complete: Corruption signatures characterized, feasibility assessed
• Phase 3 complete: Runtime model achieves measurable noise reduction
• Phase 4 complete: Practical application demonstrated OR documented infeasibility

Risks

Risk	Mitigation
Multipath makes phase unusable indoors	Start with line-of-sight outdoor tests
CFO too unstable to track	Evaluate both devices' oscillator stability
Effort exceeds benefit	Strict phase gates; stop if data looks bad
Scope creep from core therapy device	This is explicitly exploratory, not blocking
Serial bottleneck during logging	Use binary logging or decimation (1-in-10 packets)
Integer ambiguity unsolvable	Focus on velocity first, not absolute distance

---

Cross-Model Review (Gemini)

The following insights were contributed via cross-model triangulation (2025-12-23).

1. The Math is Valid: "Vernier Caliper" Analogy

Phase is just Time in a costume:

Time Domain	Frequency Domain
Clock drift (10 ppm)	Carrier rotation (3.6°/packet)
Propagation delay	Phase slope across subcarriers

The Vernier Caliper Strategy:

• FTM (ToF): The "Main Scale" — measures 5m ± 1m
• CSI (Phase): The "Sliding Scale" — measures 0.125m (wavelength) ± 0.001m
• Fusion: Use FTM to resolve which wavelength (Integer Ambiguity Resolution), CSI to locate within that wavelength

The same Kalman Filter used for UTLP timing will work for phase tracking.

2. The Trap: Integer Ambiguity (The Wrap)

atan2(Q, I) wraps at ±π. Moving 12.5cm (one wavelength) spins 360° and looks identical to the starting point.

Problem	Difficulty	Approach
Relative Motion (Velocity)	Easy	Count the spins (Doppler)
Absolute Distance	Hard	Need FTM fusion or continuous tracking

Recommendation for Phase 3: Don't solve "Absolute Distance" immediately. Focus on Phase-Coherent Velocity:

• Phase derivative = 0 → Devices are relatively static
• Phase derivative positive → Devices are closing
• Phase derivative negative → Devices are separating

This is sufficient for collision avoidance and gesture detection without solving the full integer ambiguity problem.

3. Killer App: Non-Contact Respiration Monitoring

The EMDR therapy use case Gemini identified:

Human chest movement during breathing: ~5mm

At 2.4 GHz (λ = 12.5cm): - 5mm movement = ~14° phase shift - ESP32 CSI noise floor: ~2° (when filtered) - Signal-to-noise: 7:1 — detectable!

Implication: The bilateral pulsers could sit on a table and detect patient hyperventilation by analyzing WiFi field perturbation between left and right units.

• No wires
• No cameras
• Just phase perturbation analysis

This fits the "invisible, empathetic sensing" philosophy of the mlehaptics project.

4. ESP32-C6 Technical Gotchas

A. HE-LTF (WiFi 6 Preamble)

The C6 speaks WiFi 6 (802.11ax). The config struct enables htltf (WiFi 4 High Throughput).

• If ESP-NOW uses standard rates (1 Mbps): HT-LTF is fine
• If AX rates are used: Need HE-LTF capture

Action for Phase 1: Verify rx_ctrl field in callback to confirm which preamble type is being captured.

B. Serial Bottleneck

CSI data volume:

52 subcarriers × 2 bytes (I/Q) = 104 bytes/packet
At 100 Hz packet rate = ~10 KB/s raw data

Warning: printf() over USB Serial while running the radio stack may crash the ISR.

Fix for Phase 1: - Use binary logging (SLIP or raw bytes), OR - Decimate: Log 1 of every 10 packets to "build intuition" without choking the CPU

Gemini Verdict

Approve. This is the correct next step. You have mastered Time (UTLP). Space (RFIP/CSI) is the logical next backbone. Even if you only get "Relative Motion" out of it, the ability for a swarm to know "I am moving towards my neighbor" without GPS is a game changer for the aerial drone application.

---

This ADR captures an exploration path, not a committed feature. The goal is to not forget a promising direction while remaining focused on current priorities.