Introduction to Drones, Robotics and Edge Devices

Lab A — Smart sensor node (ESP32 + MicroPython)

Story: A rural health post keeps vaccines in a fridge that must stay between 2 °C and 8 °C. Build a node that filters the temperature, detects anomalies on the device, and only sends alarms and short summaries over the network.

Wiring (DHT22)

Setup

1. Flash MicroPython on the ESP32 (download the firmware from micropython.org and follow the ESP32 instructions there; esptool is used for flashing).
2. Install Thonny (thonny.org) and select MicroPython (ESP32) as the interpreter.
3. Copy main.py to the board. Edit the Wi-Fi name, password and MQTT broker at the top.
4. To see messages, subscribe from a laptop: mosquitto_sub -h test.mosquitto.org -t "skylab/<your-group>/#" (public test broker: never send real patient data to it).

Quest checkpoints

1. Sensor online — raw readings print every 2 s.
2. Clean signal — paste your EMAFilter from Exercise 6; print raw vs filtered.
3. Watchdog — paste your detector logic from Exercise 7; breathe on the sensor or touch it with a warm finger to trigger an alarm.
4. Bandwidth saver — the node publishes only a summary every 60 s plus alarms. Count bytes sent vs a "send everything" design.

Boss challenge: run the node on battery and use machine.deepsleep() between readings. Measure how many hours a 2000 mAh power bank lasts.

Lab B — Obstacle-avoiding robot car (Arduino)

Story: A small delivery robot must move along a corridor of a health centre without hitting people or furniture.

Wiring

Quest checkpoints

1. Wheels turning — testMotors() drives forward, back, spins left and right. Relate each move to wheel_speeds_to_twist() from Exercise 3.
2. Eyes open — the serial monitor prints the ultrasonic distance in cm.
3. Don't crash — the robot drives forward and turns away when an obstacle is closer than SAFE_CM.
4. Smooth driver — replace the on/off turning with a proportional rule: the closer the obstacle, the slower the robot (same idea as go_to_goal slowing near the goal).

Boss challenge: mount the HC-SR04 on a small servo, scan left/centre/right, and always turn toward the side with the most free space (a 3-beam "lidar").

Lab C — Edge AI camera (Raspberry Pi)

Story: A camera at a farm co-operative or a clinic entrance must recognise objects locally, without uploading video to the cloud (privacy, no internet, lower cost).

Setup (Raspberry Pi OS, 64-bit)

sudo apt update && sudo apt install -y python3-picamera2
python3 -m venv --system-site-packages ~/edge && source ~/edge/bin/activate
pip install ai-edge-litert numpy pillow      # LiteRT, the new name of TensorFlow Lite
#   older OS? use:  pip install tflite-runtime

Get two versions of an image-classification model, one float32 and one int8 (quantized), plus the labels file. MobileNet models in .tflite format are published by the TensorFlow / LiteRT team (search "mobilenet tflite"), or convert your own with the TensorFlow Lite converter. Put them in this folder as:

model_float.tflite   model_int8.tflite   labels.txt

Run

python classify.py --model model_int8.tflite      # live camera
python classify.py --benchmark                    # compare float vs int8
python classify.py --image test.jpg --model model_float.tflite   # no camera

Quest checkpoints

1. First sight — the Pi prints the top-3 labels for what the camera sees.
2. Stopwatch — record inference time (ms) and frames per second.
3. Squeeze it — run --benchmark: compare file size, speed and top-1 label of the float and int8 models. Connect the result to Exercise 8.
4. Smart trigger — only save a photo when a chosen label (e.g. "person") has confidence > 0.6. Count how much storage/bandwidth that saves vs recording video.

Boss challenge: measure power draw with a USB power meter, then compute energy per inference (joules) for float vs int8.

Lab D — Programmed drone mission (Tello)

Story: Rehearse an autonomous delivery route indoors before anyone flies a real delivery drone outdoors.

The Tello is a small, light indoor drone with a Python SDK. It holds altitude and position itself (its own flight controller runs the PID loops you wrote in Exercise 1), so students command high-level moves.

Setup

pip install djitellopy

Turn on the Tello, connect the laptop to its Wi-Fi (TELLO-XXXXXX), then:

python mission.py --dry-run          # prints the plan, does not fly
python mission.py                    # flies (spotter ready!)

Quest checkpoints

1. Pre-flight check — connect, print battery and temperature; refuse to fly below 30% battery.
2. Square — take off, fly a 1 m square, land.
3. Delivery route — fly the waypoint list produced by your Exercise 9 planner (scaled down to fit the room), hovering 2 s at each "health post".
4. Telemetry log — save height, battery and attitude every 0.5 s to a CSV and plot it. Compare with the altitude plots from sims/drone_altitude_sim.py.

Boss challenge: put a printed marker (e.g. an ArUco tag) on the floor and use the Tello's camera stream with OpenCV to land on it.

Safety rules: see hardware_labs/README.md. Press Ctrl+C at any time — the script then lands the drone.