ML Detector: architecture, training pipeline, and future direction

[francescopace]

Starting from v2.5, ESPectre includes an experimental ML-based motion detector built around a simple idea: remove calibration from the user experience while keeping inference fully on-device.

Traditional statistical approaches still work very well, but calibration remains a practical limitation. A compact ML model offers a path toward a more plug-and-play detector, while preserving the project's edge-first and privacy-first nature.

At the same time, this is a constrained embedded problem. The target hardware has limited RAM, limited flash, and a very small inference budget, so the model has to stay lightweight enough for real deployment on ESP32-class devices.

The current production model is a compact MLP:

9 -> 32 -> 16 -> 1

It runs on 9 turbulence-window features extracted from CSI-derived motion signals.

This network uses 865 parameters and occupies about 3.4 KB as embedded float weights, which keeps it small enough for practical deployment on ESP32-class devices. The MLP itself requires 816 MACs per inference.

On real hardware, the full ML detection path (feature extraction + MLP inference) currently measures about:

• ~3400 us on ESP32-C3
• ~1500 us on ESP32-C5
• ~1900 us on ESP32-C6
• ~430 us on ESP32-S3

Most of the runtime cost comes from feature extraction rather than from the neural network itself.

The broader point, though, is not just the model itself. Over time, this work turned into a full ML workflow around ESPectre, including:

• labeled CSI data collection
• dataset metadata and quality controls
• training and export pipeline
• host-side live ML inspection tools
• cross-chip validation
• embedded deployment for both firmware and Micro-ESPectre

What interests me most at this stage is not simply maximizing offline accuracy, but pushing toward robust calibration-free sensing that remains practical on real hardware.

This work is also larger than motion detection alone. It is the foundation for future sensing tasks such gesture recognition and Human Activity Recognition (HAR)!

If you want the implementation details, the current references are here:

• Algorithms
• ML data collection guide
• Training pipeline
• Performance results
• Roadmap

If you are experimenting with CSI-based ML, tiny on-device models, or embedded Wi-Fi sensing pipelines in general, feel free to share results, ideas, or criticism.

[Poppinkpop]

Hi Francesco,
I am not familiar with working with WiFi Channel State Information.
My setup uses simple PIR in the living room.
However I want to show you my detection method. KIS, very KIS
The ESP works standalone and communicates notifications using Telegram.
If you're interested, contact me.
Poppink
https://github.com/Poppinkpop/ESP-motion-absence-detection