Course · BSc SEIoT year 3 · 6 chapters · 5 workshops · 5 ESP32 labs
TinyML: Embedded AI on Microcontrollers
From the concept of AI to a model running on ESP32: design approaches, memory and energy constraints, TensorFlow Lite, int8 quantisation and TensorFlow Lite for Micro — 6 chapters, 5 workshops and 5 labs.
- Faculté des Sciences de Bizerte
- 2026–2027
SEIoT 3 · Internet of Things Engineering module — lectures, in-class workshops and labs on ESP32 / NodeMCU, 2026 edition. Six steps, from the concept of AI to a model running on a microcontroller.
Objectives
By the end of the module, you will be able to:
- Explain what TinyML is and why AI is moved as close as possible to the sensor.
- Choose between hard-wired, programmed and learning-based logic according to constraints.
- Size a model: Flash memory, RAM, latency and energy.
- Optimise a model with int8 quantisation and pruning while measuring the accuracy loss.
- Deploy a TensorFlow Lite for Micro model on ESP32 / NodeMCU and integrate it into an IoT chain.
Course outline (6 chapters)
- Reflections on AI — history, vocabulary (AI, ML, DL), limits, costs and environmental impact
- Embedded systems: which design approach? — hard-wired, programmed, co-design or learning-based logic, depending on constraints
- The TinyML project — definition, microcontroller constraints, real applications, software and hardware
- TensorFlow Lite (LiteRT) — train, convert, optimise (quantisation, pruning) and run
- TensorFlow Lite for Micro — the model becomes a C array running with no OS and no malloc on a microcontroller
- Synthesis and practice — the full chain from sensor to decision, 2026 trends, final quiz and transition to labs
In-class workshops
- Deterministic or probabilistic? And how much energy?
- Choosing the right approach for six products
- Does my model fit on the board? (Flash, RAM, tensor arena)
- Measuring the effect of quantisation (Colab)
- TFLM “Hello World” on ESP32 (sine model and LED)
Labs (ESP32 / NodeMCU)
- Lab 1 — Data acquisition: ESP32, MPU6050, DHT22, serial port, Python
- Lab 2 — Gesture dataset: 4 classes, CSV, Edge Impulse
- Lab 3 — int8 classifier: deployment and latency measurement on ESP32
- Lab 4 — Anomaly detection: autoencoder on fan vibrations
- Lab 5 — Connected device: local inference + MQTT + Node-RED
Each lab builds on the previous deliverable; wiring, code, questions and grading are in the lab booklet.
Hardware and tools
- Boards: ESP32 DevKit (2 × Xtensa LX6, 240 MHz, 520 KB SRAM, 4 MB Flash, Wi-Fi + BLE) and NodeMCU ESP8266.
- Sensors: MPU6050 (accelerometer/gyroscope, I²C), DHT22, push button, LED.
- Software: Python, Google Colab, TensorFlow / Keras, LiteRT, LiteRT for Microcontrollers (TFLM), Edge Impulse, Arduino IDE / PlatformIO, MQTT and Node-RED.
Assessment and practice
- An MCQ quiz at the end of each chapter plus a final quiz, with commented answers (40 questions in total).
- 15 worked exercises: cost of AI, real-time budget, radio data volume, battery life, model memory footprint…
- 5 in-class workshops and 5 labs on ESP32.
Main references
- P. Warden, D. Situnayake — TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers, O'Reilly.
- R. Banerjee — Hands-on TinyML, BPB Online.
- G. M. Iodice — TinyML Cookbook, 2nd ed., Packt.
- A. Gupta, S. Nandyala — Deep Learning on Microcontrollers, BPB Online.
- S. Chaudhari et al. — TinyML for Edge Intelligence in IoT and LPWAN Networks, Academic Press.
Downloadable resources
Sign in or create a free account to download these files.
-
PDF
Cours TinyML_SEIoT3_v2026_QCM.pdf · 1.8 Mo
-
PDF
TPsFascicule_de_TP_TinyML_sur_ESP32_NodeMCU.pdf · 538.5 Ko