12 Machine building

Presentation Machine 1

---

Presentation Machine 2

---

Presentation Machine 3

---

👥 Mariam’s Individual Contribution — Self-Balancing Robot

Full page: Mariam’s Week 12 documentation

🤖 Main Robot

This week the group assignment required us to design a machine that includes mechanism + actuation + automation + function + user interface, build the mechanical parts and operate it manually, and document both the group project and individual contributions.

After discussing many ideas, we decided to build a self-balancing robot. Mariam also proposed adding a camera that detects a smile — and when a smile is detected, the robot deliberately loses balance 😄. This made the individual task significantly more complex, but led to exploring several interesting technologies.

---

📐 Sketches of Electronics Part and Control Logic

Here are some example sketches created during research — this approach helps build stronger understanding, functioning as a visual database from first idea to final result.

---

🔌 Main Robot — Electronics

After finishing the research, work began on the part of the robot responsible for maintaining balance. The following components are needed for the main functionality:

1. MCU — Seeed Studio XIAO RP2040

Pins D0–D7 are used as follows: D0/D1 for Motor A direction, D2/D3 for Motor B direction, D4/D5 for SDA/SCL (LCD + MPU6050), D6 for LED, and D7 (RX) to receive the signal from the face detection PCB.

2. Sensor — MPU6050

The MPU6050 is the "gold standard" for balancing projects — it combines a 3-axis accelerometer (tells the robot which way is down) and a 3-axis gyroscope (tells how fast it is falling), with 16-bit ADC for detecting very small movements.

3. Motors — 2× TT DC Gear Motor (3–6V)

This is a compact and convenient option

4. Motor Drivers — 2× A4953

The only motor driver available in the lab; its minimum voltage is 8V, which is not ideal for lower voltages.

5. Battery — 3× 1200mAh 4.2V Li-Po

Compact, easy to use, and enough to power both motors and the MCU for the camera.

6. Voltage Regulator — LM2940

Sensitive to noise, so a 22µF capacitor is used to keep the 5V rail stable.

7. Display — OLED SSD1306 (I²C)

Compact size and multiple colors allow nice animations.

8. Diode — Schottky Diode SOD123

Placed near the 5V line so that when powering from battery, USB current won't damage the MCU.

9. Resistors & Capacitors

For proper configuration and to avoid short circuits.

🗺️ Main Robot — Schematic and PCB Design

With all components and their purposes defined, the KiCad schematic and PCB designs were assembled.

3D views of the PCB:

Once the files were ready, .nc files were generated on modsproject.org for F.Cu (1/64” bit) and Edge.Cuts (1/32” bit).

---

🔧 Main Robot — Soldering

All components laid out and workspace prepared:

Final soldered result (top and back):

---

🧠 Main Robot — Programming & PID Control

The first test code initialized the MPU6050 and OLED and drove the motors forward — but without a closed-loop system the robot cannot balance. A balancing robot is like keeping a stick on your finger: without continuous correction it falls.

The robot needs PID control — a smarter brain that runs every few milliseconds:

Term	Formula	Role
🟢 P (Proportional)	Kp × error	Reacts to current tilt
🟡 I (Integral)	Ki × Σ error	Fixes long-term drift
🔵 D (Derivative)	Kd × Δerror/Δt	Reacts to how fast it is falling

motor speed = P + I + D — recalculated every few milliseconds.

Video of the first test (MPU6050 + OLED + motors, open loop):

With Hrach’s help a temporary body was built for testing. Here are the first videos of the robot falling 🤪:

After several tests, the SDA pin of the RP2040 burned out. The MCU was replaced with an ESP32-C3 (same pin functions), and Onik helped further stabilize the system.

---

😄 Smile Detection Module

Research

The XIAO ESP32-S3 Sense was chosen as a vision co-processor — it focuses entirely on the camera stream and smile detection, and communicates to the main MCU via a single GPIO pin (PIN 43: HIGH = smile detected, LOW = no smile).

Electronics — Smile Detection PCB

Components: ESP32-S3, LM2940 voltage regulator, capacitors and resistors. The PCB was also designed in KiCad.

3D views:

Soldering — Smile Detection

When attaching the ESP32-S3, double-sided tape was used under the chip to prevent its back copper layer from shorting with the PCB GND.

Final result:

Programming — Smile Detection & Web Interface

The smile detection algorithm is brightness-based (grayscale, no face detector). It compares a forehead zone (top 10–25% of frame) against a mouth zone (62–85%). If the mouth is significantly brighter (teeth visible) and local contrast is high enough, a smile is counted. A web dashboard served from the ESP32-S3 shows the live brightness chart, contrast meter, total smiles, and GPIO 43 status in real time.

Web interface screenshot:

Smile detection demo video:

Limitation: zones are hardcoded as fixed frame percentages, so detection only works when the person is centered at a specific distance. Any position change breaks the detection — the algorithm doesn’t know where the face actually is.

---

🧪 Person Classification — Edge Impulse Experiment

After the brightness-based approach proved unreliable, a custom ML model was built using Edge Impulse to classify between Mariam and Hrach.

A Python + OpenCV script auto-captured 100 face photos per person via webcam, saved into dataset/Mariam/ and dataset/Hrach/.

The dataset was uploaded to Edge Impulse → an Impulse was created (Image input → Classification) → features generated → model trained.

Model testing showed 91.36% accuracy on the test set:

The model was deployed as an Arduino Library (.zip) and installed via Sketch → Include Library → Add .ZIP Library.

Real-world result: the model recognized Hrach as Mariam, and almost any object as Hrach. The dataset of 100 images per person was far too small. 😄

---

👤 Face Detection — Final Approach

Returning to basics: the camera only needs to detect that a person is present, not recognize who. Two built-in ESP-DL models run directly on the ESP32-S3:

• HumanFaceDetectMSR01 — fast first-pass candidate detection
• HumanFaceDetectMNP01 — refines candidates with a confidence threshold

Logic: Face detected → PIN 43 HIGH → ESP32-C3 stops balancing motors. No face for 1.5 s → PIN 43 LOW → robot resumes. Green bounding boxes are drawn around detected faces and the live web interface shows camera feed + detection stats.

---

🏁 Final Robot with Interface

After many PID coefficient adjustments and code modifications, here is the final result. The robot balances and stops when a face is detected. The balancing is not yet perfect, but the full pipeline — PCB design, milling, soldering, PID control, and computer vision — was completed in one week.

---

📁 Files for Download

File	Description
PCB Design — Face Detection	KiCad PCB — ESP32-S3 face detection board
Schematic — Face Detection	KiCad schematic — ESP32-S3 board
PCB Design — Self Balancing Robot	KiCad PCB — ESP32-C3 main robot board
Schematic — Self Balancing Robot	KiCad schematic — ESP32-C3 board

---

Last update: June 17, 2026