Teach it, learn it

How a quadcopter flies

A quadcopter has no moving control surfaces. It steers only by changing motor speeds. Our drone uses an X layout, where the two diagonal pairs spin in the same direction.

• Throttle: all motors speed up or slow down together.
• Roll and pitch: motors on one side speed up while the other side slows down, tilting the drone.
• Yaw: each motor's reaction torque twists the frame the opposite way. Speeding up one diagonal pair and slowing the other turns the drone.

The F450's symmetry keeps the mixing maths simple, since every motor is the same distance from the centre.

Activity

Using the mixing explorer, predict which motors speed up for right roll, then check. Then draw the X layout and label each motor's direction from memory.

Choosing parts that work together

The motor, propeller, ESC and battery form one system. High-KV motors suit speed. Our 980 KV motors with 10-inch propellers suit stable lifting.

Here are the sums from our build:

Motor peak current ≈ 15 A → 4 motors ≈ 60 A
ESC margin        = 30 A / 15 A = 2×
Battery output    = 5.2 Ah × 60C = 312 A available
Battery voltage   = 3 cells × 3.7 V = 11.1 V nominal, 12.6 V full

We rejected a 4S battery because it would over-volt this motor and propeller pairing, and 20 A ESCs because of the risk of overheating.

Activity

Give students a different motor with a 25 A peak. Have them choose an ESC and work out the battery C rating needed for 4 motors using a 4000 mAh pack.

Power and electrical noise

The drone has two very different power needs: huge current pulses for motors and a clean, steady rail for logic. They should be kept apart.

• Battery to ESCs: thick wires through the power board.
• Battery to 5 V: a switching buck converter, which wastes far less heat than a linear regulator.
• 5 V to 3.3 V: small onboard regulators.

When motors spin up, the rail dips and rings. On our drone that froze the I2C bus. Capacitors placed right at each sensor supplied local current and absorbed the noise.

Read the full story.

Activity

With an oscilloscope, view a 3.3 V sensor rail while a noisy load switches. Add a ceramic and an electrolytic capacitor at the sensor and compare. No propellers are needed.

Sensors and buses

Reading the IMU over I2C limited our control loop, so we moved it to SPI. The magnetometer stayed on I2C because it updates more slowly. But I2C can lock up, so the firmware has a recovery routine: it toggles the clock 9 times to release the bus.

Activity

Work out how long a 14-byte IMU read takes at 400 kHz I2C and at 10 MHz SPI, ignoring overhead. Then compare both with the 4000 µs loop period.

Sensor fusion

A gyroscope is smooth but drifts over time. An accelerometer knows where gravity points but is badly affected by vibration. Sensor fusion uses each to correct the other.

1. Filter: a hardware filter of about 41 Hz, then a software filter.
2. Correct: remove false acceleration when the IMU sits away from the centre of rotation.
3. Compare: the Mahony filter compares gravity as predicted from the gyro with gravity as measured, using a cross product.
4. Adjust: a small PI loop nudges the gyro to remove the error and learn its drift.
5. Yaw: gravity cannot sense heading, so yaw blends 98% gyro with 2% magnetometer.

Orientation is stored as a quaternion, which avoids the gimbal lock of Euler angles.

Activity

Log roll angle from the dashboard while tilting the board by hand. Then tap the frame and watch how little the filtered value reacts.

PID control

PID turns an error into a correction. The firmware runs it 250 times a second.

• P pushes in proportion to the error now. Too much makes fast oscillation.
• I adds up small errors over time. It removes steady offsets, but too much makes slow rocking.
• D resists fast change, like a shock absorber. Too much amplifies noise and heats motors.

Our firmware adds refinements to the textbook version:

• D on the measurement, to prevent a kick when sticks move suddenly.
• A low-pass filter on the D term.
• Feedforward for a sharper stick response.
• Pausing the I term when outputs are saturated, to prevent windup.
• Shortest-path yaw error, so 350° to 10° is a 20° turn.
• TPA, which reduces P and D at high throttle.

Activity

On the simulator, set I and D to 0 and raise P until it oscillates. Add D until it settles. Turn on the steady push, see the offset, then add I. Later, repeat on a rig with a real axis.

Real-time firmware and failsafes

A flight controller must run on time, every time. Our main loop is driven by a hardware timer:

TIM4 interrupt every 4000 µs → sets pid_loop_flag
while (1):
  if pid_loop_flag:
     read IMU → fusion → PID → motor mix → set PWM   (250 Hz)
     every 5th loop: barometer + magnetometer        (50 Hz)
     every 50th loop: send telemetry JSON            (5 Hz)

Nothing here should block. UARTs use interrupts with lock flags, the radio is polled once per loop, and hardware timers generate PWM in the background. Gains are stored in internal flash, with a magic number that detects blank memory on first boot.

Failsafes cover radio loss, dashboard loss, geofence breach, crashes and I2C faults. They need testing, and they can be wrong: our low-battery failsafe is currently inactive because the battery reading is hardcoded. See the status list.

Activity

With propellers off and the drone armed, switch off the transmitter and record how long it takes to react. It should be about 500 ms. Discuss why a failsafe needs testing. As an extension, propose a fix for the battery reading.

Radio, telemetry and dashboard

The drone has two separate links:

Control:   Joysticks → ADS1115 → ESP32 → nRF24 ))) nRF24 → STM32        50 Hz, 8 bytes
Telemetry: STM32 → UART → ESP32 ))) Wi-Fi → Node.js → Socket.IO → browser  5 Hz, JSON

The control packet has four 16-bit values, packed with #pragma pack so both chips agree on its exact layout. A low 250 kbps data rate gives better receiver sensitivity and more range, and 8 bytes needs very little bandwidth.

The dashboard limits its heavy chart updates to 1 Hz, so the page stays responsive.

Activity

Pick one value, such as roll angle. In the source code, follow it from main.c to ESP32_Telemetry.ino, server.js and script.js. Draw the path.

A magnetometer as a metal detector

Earth's magnetic field is fairly uniform locally. Ferrous objects distort it. Fly a magnetometer over an area and tag each reading with GPS, and those distortions appear on a map.

The dashboard's survey view cleans the data. It rejects poor GPS and readings taken while turning, corrects for altitude and slow drift, and draws heatmaps.

Limits: the sensor is a consumer compass chip and GPS error is 2 to 5 m. It can find a zone, but not an exact point, depth or shape. A signal can also come from the drone itself, which is why mounting and calibration matter.

Activity

Lay out a 1 m grid on the floor. Move a magnetometer board by hand over a hidden steel object and record the reading at each point. Plot it and locate the object. Repeat at twice the height.

Running this as a course

The lessons suit university or advanced school students with basic programming and circuits. In total they take about ten hours, plus hands-on rig work if you have a drone. Most activities need no flight at all.

Suggested schedule

Marking rubric