Commodity ESP32 WiFi Sensors: Ranging & Gesture Use Cases

• Commodity ESP32 WiFi sensors are off-the-shelf, low-cost modules that utilize IEEE 802.11mc FTM and CSI extraction to enable indoor ranging, motion detection, and gesture tracking.
• Advanced signal processing and machine learning methods, such as regression trees and Gaussian processes, are employed to reduce ranging errors by 10–20% compared to raw RTT measurements.
• Despite their cost and ease of deployment, these sensors face limitations like coarse spatial resolution and multi-user separability challenges, necessitating sensor fusion and calibration for robust performance.

Commodity ESP32 WiFi sensors refer to off-the-shelf, low-cost wireless modules—primarily the Espressif ESP32 and derivatives—used for precision sensing and indoor localization tasks via standard WiFi protocols. Leveraging capabilities such as Fine Time Measurement (FTM) per IEEE 802.11mc and Channel State Information (CSI) extraction, these devices enable ranging, motion detection, gesture tracking, and activity monitoring, with computational and deployment characteristics determined by their cost-optimized hardware design and open-source firmware. Despite their accessibility and integration with edge devices, their physical and protocol limitations impose constraints on spatial resolution, multi-user separability, and robustness under multipath-rich or heterogeneous environments.

1. IEEE 802.11mc FTM and Ranging Theory

The foundation for accurate RF-based ranging in ESP32-XX series modules is the IEEE 802.11mc FTM protocol, which enables two-way ranging between an initiator (Tag) and a responder (Anchor). The operation involves an exchange of time-stamped frames:

• Tag transmits FTM_REQ at timestamp $T_1$.
• Anchor receives at $T_2$, delays, and responds at $T_3$.
• Tag receives FTM_RESP at $T_4$.

Timestamps are 64-bit counters (1 ps resolution). The raw round-trip time (RTT) is

$\text{RTT}_\text{raw} = (T_4 - T_1) - (T_3 - T_2),$

from which the one-way time-of-flight (ToF) is $\text{ToF} = \text{RTT}_\text{raw}/2$. The distance is

$d_\text{raw} = (\text{RTT}_\text{raw} \cdot c)/2,$

where $c \approx 3 \times 10^8$ m/s, or equivalently $$d_\text{raw}[\text{m}] = (\text{RTT}_\text{raw}[\text{ns}] \times 0.15)$$ (Vales et al., 2024). The ESP32’s internal firmware implements additional linear compensation with breakpoints, tuned primarily for indoor RSSI levels.

2. ESP32 Firmware, Hardware, and Sensing Pipeline

ESP32 modules supporting FTM require ESP-IDF v4.3 or later. The responder is enabled via:

wifi_ftm_responder_config_t resp_cfg = { .enable = true, .burst_period = 0 }; esp_wifi_ftm_responder_enable(&resp_cfg);

The initiator configures the FTM session with desired channel, bandwidth (HT20 or HT40 for 2.4 GHz), and number of frames:

wifi_ftm_initiator_config_t init_cfg = { .burst_period = 0, .num_frames = 8, .channel = 1, .bandwidth = WIFI_BW_HT40, .retries = 1 }; esp_wifi_ftm_initiator_start(ap_bssid, &init_cfg);

RTT and RSSI results are delivered by callbacks registered to WIFI_EVENT_FTM_REPORT.

ESP32s are typically equipped with single or dual PCB trace antennas, up to 40 MHz bandwidth (S2/C3: 2.4 GHz only), and exhibit TX peak currents up to 450 mA during FTM rounds. Idle and deep-sleep consumption drops to 73 mA (TX idle) and 23 μA, respectively.

For CSI extraction, the “esp-wifi-csi” driver provides per-packet amplitude and phase per subcarrier (e.g., 64 subcarriers for 20 MHz BW), though phase is generally unreliable due to offset impairments (Zhu et al., 4 Jun 2025).

3. Practical Performance and Empirical Findings

A. Ranging and Localization

Indoor experiments (e.g., 4 anchors in a 12×6 m room, HT40) show 75% of FTM ranges within 5 m of the ground truth, but NLOS and multipath can introduce errors up to 20 m. Raw RTT-derived distances perform ~50% worse than the firmware’s “dist_est,” but both are limited by first-path detection, multipath bias, and compensation tuned for typical RSSI (Vales et al., 2024).

Outdoor (open-air) setups achieve 90% of measurements with <1.5 m error (BW=40 MHz) and <2.5 m error (BW=20 MHz), with performance primarily limited by clock drift and processing delay estimation. The theoretical minimum error is bounded by channel bandwidth and ToF quantization: for 40 MHz, raw resolution ~1.5 m.

B. Motion and Gesture Sensing

Commodity ESP32 modules, via CSI monitoring, support wide-area human motion sensing and basic gesture detection. The amplitude cohort extracted over 14–64 subcarriers enables detection of human gait and motion through walls. Promiscuous mode and the CSI driver permit per-packet collection at 100–300 Hz, with denoising via median filters and optional low-pass postprocessing. Motion is declared when the global motion statistic (derived from subcarrier autocorrelation) exceeds the empirically determined threshold ($\eta\approx0.15$–$0.25$) (Zhu et al., 4 Jun 2025).

Deployed across 10 million+ commercial devices, similar statistical pipelines have produced human motion detection accuracy of 92.6% (across 4 million samples and 280 devices); false alarms due to pets or non-human triggers are reduced from 63% to 8.4% using feature-driven SVM filtering with temporal confidence fusion.

C. Multi-User Limitations

Extensive trials with multi-person gait identification (1–10 individuals) using ESP32s with 3 antennas and 52 usable subcarriers consistently result in classification accuracy of 39–56%, irrespective of the blind-source-separation algorithm (FastICA, SOBI, PCA, NMF, Wavelet, Tensor). Diagnostic metrics show intra-subject variability (ISV) and inter-subject distinguishability (ISD) are dominated by hardware-induced noise and low angular resolution ($\Delta\theta\approx67^\circ$ for $k=3$ antennas), with >97% class feature overlap. Non-monotonic performance degradation rates (PDRs) and environment-dependent robustness underscore fundamental limitations imposed by commodity chipsets (Custance et al., 5 Jan 2026).

4. Advanced Signal Processing and Machine Learning Approaches

Machine learning approaches augment raw ESP32 ranging by mapping observed RTT and mean RSSI to corrected range estimates:

• Regression Tree (min leaf size 4), SVM (RBF kernel), Gaussian Process (exp kernel), and shallow neural nets (1 layer, 100 neurons) were evaluated on ~7000 samples over diverse indoor/outdoor environments.
• In cross-scenario validation (completely new environment), ML correction reduced absolute error by 10–20% beyond chip “dist_est.” Regression trees and Gaussian Processes achieved best robustness, with regression tree models small enough for real-time deployment (~403 KB flash).
• Data pipeline for edge deployment: on-board feature normalization, prediction, and reporting (UART/ROS) incur negligible latency or power cost ($<1$ μA extra sleep draw) (Vales et al., 2024).

In dense deployments, feature sets for SVM-based motion classification include stride parameters, speed statistics, and autocorrelation metrics. Edge-optimized implementations eschew floating-point arithmetic, deploy Q15 fixed-point, and selectively transmit compact feature vectors on event detection (~2 kB per 6 s window), slashing uplink by 99.7% vs. raw CSI.

5. Application Prototyping and Deployment Guidance

A. Calibration, Co-location, and Coverage

Open-air per-anchor calibration is essential for bias correction. Calibration for elevation becomes critical in 3D layouts or when anchors are placed on different floors. In trilateration contexts, FTM ranges should be fused with RSSI, Time Difference of Arrival (TDoA), or Angle of Arrival (AoA) data for improved spatial robustness (Vales et al., 2024).

The optimal physical separation between sensor ("Bot") and AP ("Origin") is below 6.5 m; coverage beyond 10 m suffers SNR degradation. Distributed deployment (multiple Bots/APs) allows zone-based proximity differentiation and, when motion is concurrent, yields >90% correct multi-occupant separation for proximity (but not identity).

B. Power Management and Communication Considerations

Power consumption for FTM operations peaks at 450 mA (TX active), with low-power operation accomplished via deep sleep. Duty-cycled operation (e.g., 1 range/min) enables 2–4 month unattended battery life on conventional cells. Edge–cloud partitioning and event-driven MQTT or CoAP transmission minimize bandwidth and avoid contention with standard WiFi traffic.

C. Environmental Constraints

LOS conditions are paramount for sub-meter accuracy. Multipath and NLOS (e.g., glass, metal partitions) introduce positive bias (overestimation) with heavy-tailed error distributions. Indoor sub-meter resolution is not attainable with ESP32 WiFi alone; fusion with UWB or inertial sensors is required for higher precision (Vales et al., 2024).

6. Passive Gesture Tracking and Path Reconstruction

Centimeter-level passive gesture tracking has been realized using three ESP32s (one TX, two RXs in promiscuous CSI capture mode), with all signal processing performed on a host PC. The real-time pipeline includes:

1. CSI denoising via complex ratio between antennas, removing amplitude and phase artifacts from AGC and SFO/CFO.
2. PCA to extract principal components corresponding to motion-induced reflections.
3. Phase unwrapping to convert angle changes to path-length variations: $\Delta d(t) = (\lambda/2\pi) \Delta \phi(t)$.
4. MUSIC algorithm for AoA estimation, with MDL-driven model selection, and triangulation for 2D hand positioning.
5. Post-processing via static component elimination algorithm, segmenting the complex trace and recalibrating for drift to yield $\Delta d_\text{corrected}(t)$.

This approach achieves sub-2 cm median 2D tracking error over 1.5 m, with end-to-end latency below 20 ms (Han et al., 2020).

7. Practical Limitations and Research Directions

Commodity ESP32 WiFi sensors, while effective for single-person ranging, motion and coarse gesture detection, are fundamentally limited for high-resolution, multi-person activity analysis by restricted antenna count, subcarrier number, and SNR. Future improvements require moving to massive MIMO (≥8 antennas), mmWave, or hybrid sensor modalities. Diagnostic metrics such as ISV, ISD, and PDR should be adopted routinely in early-stage prototyping to validate the feasibility of WiFi-CSI-based inference in deployment contexts (Custance et al., 5 Jan 2026). Persistent challenges remain in real-time calibration adaptation, dynamic machine learning updating, and robust fusion across diverse hardware and environments. Open-source measurement sets and firmware libraries facilitate transparent benchmarking and extension of these distributed sensing capabilities.