Element 10.11
Machine learning on microcontrollers (TinyML)
- Sheet
- 10.11
- State
- Drafted
- Release
- 16 November 2026
- Editor
- Contributed by Praecise as open research
- Reviewers
- To be named at review
- Safety
- S1, as the chapter
- Licence
- CC BY-SA 4.0
- Resources
- 6
The element
- Requirement
Where a device must recognise a pattern (a sound, a movement, an image, an anomaly) without sending raw data away, it runs a small trained model on the microcontroller, measured for accuracy and speed on the hardware it will use.
- How to tell
- The model runs within the memory of the chip. TensorFlow Lite for Microcontrollers targets devices with tens to hundreds of kilobytes of memory and is tested on Arm Cortex-M, Cadence Xtensa and Qualcomm Hexagon, with Arduino and ESP32 as community platforms (TensorFlow Lite for Microcontrollers).
- Accuracy and latency are measured on the device against a stated target. MLPerf Tiny gives reference tasks and quality targets: keyword spotting at 90 per cent top-1, person detection at 80 per cent, image classification at 85 per cent, anomaly detection at 0.85 AUC, and streaming wake-word with no more than eight false positives and eight false negatives (MLPerf Inference: Tiny).
- Energy per inference is measured where battery life matters; MLPerf Tiny includes optional energy measurement (MLPerf Inference: Tiny).
- Raw data that could identify people stays on the device unless there is a reason and consent to send it.
- Options
- A sensor board with an on-board hub that already runs activity detection: the Nicla Sense ME (Arduino Nicla Sense ME).
- A general microcontroller running TensorFlow Lite for Microcontrollers (TensorFlow Lite for Microcontrollers).
- A larger board for vision: the Portenta H7 with its vision shield runs models on live video (Arduino Portenta H7).
- Send data to a computer and run the model there: right when the device has power and a link and privacy allows.
- Settled, open, contested
Settled: a common benchmark with stated tasks and targets exists for this class of device (MLPerf Inference: Tiny), and an open runtime under Apache-2.0 (TensorFlow Lite for Microcontrollers). Open: accuracy in the field. Benchmarks use fixed datasets; no open source opened here measures how far community-collected data moves the result. Contested: on device or on a server. On-device models protect privacy and save radio energy. Server models are larger and easier to update. The choice follows power, link and the sensitivity of the data.
The four stages
- Learn: the reference tasks (MLPerf Inference: Tiny); the runtime's examples (TensorFlow Lite for Microcontrollers).
- Design: pick the task and the target accuracy; check memory against the chip (RP2350, ESP32-C6).
- Build: train on data from the place, convert, and run with TensorFlow Lite for Microcontrollers.
- Deploy: measure accuracy in the field and keep the model version with the firmware version (10.4).
Resources at each stage
1 Learn
Understand
- Arduino Nicla Sense ME Tool, S0
- Arduino Portenta H7 Tool, S0
- ESP32-C6 Tool, S0
- MLPerf Inference: Tiny Standard, S0
- RP2350 Tool, S0
2 Design
Size and choose
- Arduino Nicla Sense ME Tool, S0
- Arduino Portenta H7 Tool, S0
- ESP32-C6 Tool, S0
- MLPerf Inference: Tiny Standard, S0
- RP2350 Tool, S0
- TensorFlow Lite for Microcontrollers Library, S0
3 Build
Build or obtain
- Arduino Nicla Sense ME Tool, S0
- TensorFlow Lite for Microcontrollers Library, S0
What changes between 30 adults and 300 people
At 30 adults, one ready model, such as wake-word or activity detection, on a sensor board. At 300 people, models trained on the settlement's own data for its own sounds, animals or machines, with a person who tests each model before it goes out.
Reference designs
Open for contributions Designs with drawings, specifications and a bill of materials. Add to it
Parts list
Every resource that supports this element. Those marked External carry the label “External resource. Not made or verified by Public Protocol”.
| Resource | Type | Origin | Maturity | Safety | Licence, as stated at its source |
|---|---|---|---|---|---|
| Arduino Nicla Sense ME | Tool | External | Established | S0 | Not stated on the page opened |
| Arduino Portenta H7 | Tool | External | Established | S0 | Not stated on the page opened |
| ESP32-C6 | Tool | External | Established | S0 | All rights reserved (page copyright notice) |
| MLPerf Inference: Tiny | Standard | External | Established | S0 | Not stated on the page opened |
| RP2350 | Tool | External | Working | S0 | Not stated on the page opened |
| TensorFlow Lite for Microcontrollers | Library | External | Established | S0 | Apache-2.0 |
Depends on, and completes
Depends on 10.4.
It is one of 11 elements that complete Electronics and circuit design.
The standing questions
Who collects and labels the training data. Who checks a model before it is trusted. Who decides what may be recorded.
Environments
No source found for any setting (islands and coasts, arctic cold, hot deserts, tropical humid forest, temperate forest, high altitude, flood plains) on how on-device models change by environment.
Safety
State and evidence
Drafted
An editor has written the element's text from the resources listed under it, citing them. It is checked when it is reviewed.
Evidence from a real place or system is recorded here, with what did not work. What the states mean
History
- Drafted in the release of 16 November 2026.