Project Bee · master plan · written 24 Oct 2026

From zero to a 4 cm flying camera

Six versions, each one teaching you the next layer. Every idea is explained in plain words first, maths second. Every number is either checked against a source or clearly marked as an estimate for you to replace with your own measurement.

7 stagesv0 bench → v6 swarm

≈ 32 weeksat 35+ hours a week

≈ week 6first free flight (v1)

₹30k+cheapest path to v6 (§7)

0Start here

This page is your project's home base. It's a plain file in your project folder, with no internet needed except for links. Your ticked boxes in the plan are saved in this browser.

Today: read sections 1–4 (about 40 minutes). They tell you what we're building, why in this order, and what's possible.
Every day: open §6 The build plan. Work through 3-day blocks and tick them off. Put your real start date in and every date updates.
When a block says "read §5.x": read that part of §5. It's the textbook, so don't try to read it all upfront.
When a block says "order": go to §7.

The coloured boxes mean different things:

In plain wordsThe idea with no jargon, usually with an everyday comparison.

The mathsThe real equations, for when you want to go deeper. You're a Maths & Computing student, so this is where you'll beat most hobbyists.

Watch outThings that commonly break, burn or crash.

Checked factVerified against a datasheet, manufacturer page or the law text (sources in §10).

Numbers with "≈" are estimates. Things marked weigh it have no published figure. Your 0.01 g scale (first order, §7) becomes the source of truth. Hardware always surprises you. When something here turns out wrong, tell me and we'll update the plan.

1What we agreed

Your answers, and what each one changes in the plan:

Topic	You said	What it means for the plan
Size	Two stages: learn at ~6.5 cm, then true 4 cm	v1–v3 are 7–8.5 cm overall, where flying is forgiving. v5 "Queen" targets a 40 × 40 mm box including propellers. Your original limit was "4 or 5 cm", so up to 50 mm is acceptable if the physics demands it.
Where it flies	Indoors → calm outdoors → high outdoors on a map	v1–v3 indoors, v4 "Forager" outdoors. It's bigger, because wind beats tiny drones (§5.1).
Camera	Live view, SD recording, FPV goggles, easiest first	v2: live view + SD recording. v4: optional analog FPV (near-zero delay). v5: camera at bee size.
Control	Phone, laptop + gamepad, RC radio, own remote	v1: phone web page + laptop gamepad. v3: your own ESP-NOW remote. v4: ExpressLRS RC radio, where long range matters. Moving the RC radio to v4 is a budget choice (§7).
Flight code	Your own, plus a safety net	You write the flight controller with me, and you'll understand every module's logic. An open-source firmware (ESP-Drone family) stays flashable so you can always ask "is it my code or my hardware?"
Features	Everything, by the last version	All of it is in the ladder (§3). Honest physics: a 4 cm drone can't carry everything at once, so the Queen uses swappable "backpacks" (flow sensor / GPS / payload), like Bitcraze's decks.
Budget	₹15k–35k	The cheapest path to v6 is ≈ ₹30–35k. A comfortable path is ≈ ₹46–52k. §7 shows exactly where the money goes and what to cut.
Time	35+ hours a week	≈ 3 weeks of bench + simulator, first free flight ≈ week 6, v6 ≈ 32 weeks from start.
Skills	Starting from zero	v0 starts with fundamentals. Every firmware module gets a "logic card" (§5.10) so you understand what and why without memorising syntax.
Phone	iPhone and Android both	Phone control is a web page served by the drone, so it works on both with no app store. Android is also used for the cheap analog-FPV USB receiver.
Look	Function first, bee later	Bare boards while learning. A printed bee shell arrives in v5.
Explanations	Plain words first, then maths	That's how all of §5 is written.

2Gemini's plan, reviewed

Gemini got the direction right: a quadcopter (not flapping wings), an ESP32-S3 brain, and heavy AI on the laptop. Several specifics were wrong or risky, though. Here's each claim checked:

Gemini said	Verdict	Why	What we do instead
Flapping wings can't carry a battery + camera at 4–5 cm. Build a quad.	Right	Correct. Even Harvard's RoboBee flies tethered.	Quadcopter.
A 45–48 mm PCB with 716 motors and 31 mm props is a "4–5 cm" drone.	Wrong	The props stick out past the board. Motor spacing ≈ 41 mm + one 31 mm prop ≈ 72 mm overall. Geometry fact: a square box of side B fits four non-overlapping props of diameter at most B/2 (§5.1). A real 4 cm drone has ≤ 20 mm props.	Be honest about size. v1 ≈ 7 cm overall. v5 is designed for the 4 cm box from scratch.
Total weight 17–19 g with the custom PCB	Optimistic	Seeed's own ESP-FLY (XIAO ESP32-S3, 46 mm frame, 615 motors) weighs 25 g with its 250 mAh battery, and that's before a camera or position sensors.	Weight calculator (§5.1) + weigh every part.
XIAO + camera ≈ 2.5 g	Unverified	Seeed publishes no weight, and retail listings include packaging.	Weigh it the day it arrives.
MPU6050 IMU	Partly	DigiKey lists the MPU-6050 as obsolete, and cheap clones vary. But the safety-net firmware supports it out of the box, which helps a lot early on.	Use it for bench + v1–v4 breakouts. From v5's custom board: ICM-42688-P on SPI (the IMU used in current FPV flight controllers).
Pin plan: D8 = magnet, D9 = battery sensing	Wrong	On the XIAO ESP32S3 Sense, D8/D9/D10 are the microSD card's wires. That plan breaks recording. Also, the XIAO can't read its own battery voltage, so you need a divider on an ADC-capable pin.	Proper pin plans per version (§5.7).
SI2302 MOSFETs	Works	OK, but the AO3400A is better: 48 mΩ vs 75 mΩ at 2.5 V gate, and a 5.7 A vs ~2.7 A rating. Stalled motors have killed Si2302s.	AO3400A.
470 µF electrolytic capacitor	Heavy	It works, but it's big and tall for a micro.	Ceramic capacitors + low-battery landing in firmware. On custom boards: a buck-boost regulator for rock-solid 3.3 V.
One angle-PID loop	Simplified	Real flight controllers use two loops: angle → rotation-rate (gyro). The inner rate loop is what makes them feel locked-in.	Cascaded loops from v1 (§5.5).
ORB-SLAM3 visual mapping from the stream	Very hard	SLAM needs clean, well-timed frames. A compressed, laggy Wi-Fi stream from a tiny sensor is the opposite.	Markers first (robust), optical flow onboard, SLAM as a stretch experiment.
LionCircuits: ₹799 for 5 PCBs, 5–6 days	Right	Checked against their site.	We use them for hand-solderable boards.
Complete 6-week timeline	Unrealistic	Too fast for someone starting from zero, and it's moot now there's no deadline.	≈ 32 weeks, with checkpoints.
OV2640 camera	Outdated	Discontinued. Seeed now ships the Sense with an OV3660 or OV5640. The same driver supports all of them.	Use whatever arrives.
(missing)	Gaps	No simulator (your biggest advantage), no thrust testing before choosing motors, no flight logging, no LiPo safety, no Indian drone law, no "learn to fly".	All added.

3The big picture

In plain words

A quadcopter is four fans and a very fast reflex. It can't stay level by itself, and no two motors are identical, so it would flip in under a second. A tiny computer checks its tilt about 500 times a second and nudges the four fan speeds to correct it. That reflex is the whole trick. Everything else (camera, GPS, AI, maps) sits on top of it.

The version ladder

Each version adds one hard new thing. You never change the size, the electronics and the software all at once. That's how you always know what broke.

Version	Size (overall)	Flies	New hardware	New abilities	≈ Weeks
v0 Egg	— (desk)	Nowhere	XIAO ESP32S3, IMU breakout, one motor + MOSFET, thrust stand	Your simulator, reading sensors, knowing "which way is up", spinning a motor, measuring thrust, designing your first PCB	3
v1 Larva	≈ 7 cm	Indoors	Your first custom PCB (hand-soldered), 4 coreless motors, 1S LiPos	Your own flight code, stable manual flight, phone + gamepad control, telemetry, failsafes	4
v2 Worker	≈ 7–8.5 cm	Indoors	PCB #2, camera (Sense), height laser, optical-flow sensor, payload magnet	Hovers by itself, live video to phone/laptop, SD recording, grab & drop	5
v3 Scout	same as v2	Indoors	Printed markers, your own remote (ESP32 + joysticks)	Knows its 3D position in the room, flies routes, follow-me AI, gestures & voice, perch & watch, room map, Kalman filter	4.5
v4 Forager	≈ 12–14 cm	Calm outdoors → high	PCB #3, bigger motors, barometer, GPS + compass, ExpressLRS RC radio, optional FPV camera	Altitude by air pressure, GPS position on a map, tap-to-fly missions, return-home, geofence, long-range control	5.5
v5 Queen	4 cm (≤ 5)	Indoors + calm hops	Chip-level custom board (factory assembled), tiny motors, swappable backpacks	Everything from v2–v3 at bee size, a GPS backpack for calm days, bee shell, on-board AI experiment	7
v6 Hive	2–3 Queens + dock	Indoors	Extra Queens (same PCB batch), landing/charging dock	Swarm formation flight, auto-dock & recharge, full mission loop, open-source release	4

Why the first versions are bigger: at 7–8 cm you can carry a camera, sensors and mistakes. The Queen gets designed once everything is proven.

Why this order?

v0 before anything flies: the simulator and bench tests catch about 80% of first-flight disasters (wrong motor direction, wrong sign in a formula, a noisy sensor) on the desk, where they cost nothing.
v1–v3 at an easy size: all the software (estimator, controllers, autonomy, AI) gets written and proven where there's spare thrust.
v4 bigger, outdoors: wind doesn't care about your dreams. A 15 g drone can't fight a breeze, but a ~50 g one can handle calm days. GPS and map work gets learned here.
v5 shrinks proven software: only the hardware is new, and it's the hardest hardware (chip-level, factory-assembled).
v6 multiplies it: the same PCB batch from v5 gives you the swarm.

What's inside (v2 shown)

The thick orange line is motor power: several amps, straight from the battery to the motors, never through the brain. Two cores, two jobs: the flight reflex must never wait for Wi-Fi or the camera. That's the single most important design rule in the firmware.

Your timeline

Computed from the start date in §6 (default 25 Oct 2026). Each segment = one version. Expect slips: parts arrive late and things break. That's normal.

4Everything it could do

You asked for every possibility, so here's the full menu. The chip shows which version it lands in. "stretch" means a bonus experiment once the core works.

Flying

Self-levelling: let go of the sticks and it levels out v1
Holds its height, auto take-off & landing v2
Hands-off hover (holds position) v2
Flips & "acro" mode, just for fun v2
Throw-to-launch: toss it and it catches itself mid-air v3
"Headless" mode: sticks move it relative to you, not its nose v3

Seeing

Live video to phone and laptop v2
Record video / snap photos to microSD v2
Zero-lag analog FPV with goggles or an Android receiver v4
Face detection running on the drone itself (Espressif ESP-WHO runs on the S3) v5
"Heat vision" backpack (tiny thermal sensor) to spot warm bodies stretch

Knowing where it is

Height by laser (indoors) and air pressure (outdoors) v2 v4
Speed over the floor with optical flow v2
Exact 3D room position from printed markers v3
Draws a map of your room from its flights v3
GPS position + compass heading on a real map v4
Visual SLAM: mapping without markers stretch
UWB / "Lighthouse" positioning for centimetre accuracy indoors stretch

Doing things by itself

Flies routes you draw on the room map v3 or GPS map v4
Follow-me: a person, face or coloured ball v3
Precision landing on a marked pad v3
Perch & watch: lands on a ledge, motors off, keeps streaming v3
Return-home, invisible fence, lands itself on low battery v4
Stops before walls (tiny forward laser) v3

Talking to it

Phone web-page joystick (iPhone and Android) v1
Laptop + gamepad ground station v1
Your own handheld remote v3
ExpressLRS RC radio, with battery and GPS shown on the radio's screen v4
Hand gestures and voice commands v3
"Land on my palm" v5

Carrying things

Magnet grab & drop (1–5 g on v2, ~10 g on v4, ~1–2 g on the Queen) v2
Plant a tiny Bluetooth beacon somewhere and fly away v5
Light painting: LEDs + long-exposure photos make art in the air v2
"Bee nose": an air-quality sensor that maps a room's air stretch

Teams

Swarm formation flight (2–3 Queens) v6
Hive dock: lands on metal contacts and recharges itself v6
"Waggle dance": one bee finds a target and tells the others where v6
Radio relay: one bee extends another's range stretch

Maths & Computing specials

Extended Kalman filter fusing every sensor v3
System identification: measure your drone's real physics and make the sim match v1
Digital twin: replay any flight in 3D from its log v3
Best "flower tour": visit N markers in the shortest order (travelling salesman) v3
Model Predictive Control: a controller that plans ahead stretch
Reinforcement learning: train a controller in your sim, then fly it for real stretch

Honest limits

Wind: the Queen is an indoor creature. The Forager handles calm days, not breezy ones.
Flight time: ≈ 3–4 min for the Queen, ≈ 5 min for v1/v2, ≈ 8–10 min for the Forager (estimates).
Video delay: Wi-Fi video is 0.1–0.3 s behind reality. That's fine for watching and for AI, but never for keeping the drone level.
Range: Wi-Fi ≈ tens of metres. ExpressLRS control reaches much further, but you must keep the drone in sight and inside the law (§8).

5How it works inside

This is the textbook. Each part starts in plain words, then gives the maths. The build plan tells you when to read each one.

5.1 Why it flies, and why tiny is hard

In plain words

A propeller is a spinning wing that throws air downward, and the air pushes back up. Spin it twice as fast and you get about four times the push. The drone hovers when the four pushes add up exactly to its weight.

Moving a lot of air slowly is cheap. Moving a little air fast is expensive. Big props hover efficiently. Tiny props burn the battery fast. That one fact explains why tiny drones fly for only 3–5 minutes.

The maths

Thrust grows with the square of rotor speed, and hover means total thrust = weight:

T_{i} = k_{T} ω_{i}^{2} \sum_{i = 1}^{4} T_{i} = m g

Momentum theory (the classic helicopter result) gives the minimum power to hover. A is the total propeller disc area, ρ ≈ 1.2 kg/m³ is air density, and η (efficiency, ≈ 0.18–0.3 for tiny brushed drones, an estimate) covers motor and prop losses:

P_{ideal} = \frac{T^{3 / 2}}{\sqrt{2 ρ A}} P_{real} \approx \frac{P_{ideal}}{η}

Power ∝ 1/√A ∝ 1/D for the same weight. Going from 31 mm props to 20 mm props costs 31/20 ≈ 1.55× more power. That's the price of tiny.

The 4 cm box rule

Put four props in a square box of side B. Each prop must stay inside the box and not touch its neighbours.

The maths

Motor centres at (±s, ±s), prop radius r:

s + r \leq \frac{B}{2}, s \geq r \Rightarrow r \leq \frac{B}{4} \Rightarrow D \leq \frac{B}{2}

A 40 mm box allows props of ≤ 20 mm. A 50 mm box allows ≤ 25 mm, with 56% more disc area (≈ 25% less hover power).

Shopping trapA "65 mm whoop frame" means 65 mm between diagonal motors (the wheelbase), not its overall size. With props and ducts it's about 9 cm across. Many "4 cm drones" quote the body without the props, too.

Thrust-to-weight and hover throttle

In plain wordsIf the motors can lift twice the drone's weight (a 2:1 ratio), the drone hovers at about 70% power and keeps 30% spare to correct tilts and fight gusts. Below about 1.5:1 it feels drunk. Aim for 2:1 or better.

The maths

Thrust ∝ ω², and on a brushed motor ω is roughly proportional to throttle u (0–1). So:

T \approx T_{max} u^{2} \Rightarrow u_{hover} \approx \sqrt{\frac{W}{T_{max}}}

At 2:1 that's ≈ 71%, and at 3:1 ≈ 58%. Your thrust-stand test in block 5 will show how true this is for your motors.

Why wind beats small dronesWind force grows with frontal area (size²), but what the drone can push back with grows with its weight (≈ size³). Halve the size and the wind gets twice as strong relative to the drone. This is the square–cube law, and it's why the outdoor version is the big one.

Weight, thrust & flight-time calculator

Pick a version to load starting guesses, then type your real measured grams as parts arrive. The model is rough physics (momentum theory with an efficiency guess), so trust it for comparing designs, not to the second.

Part	grams

Max thrust per motor (g)
Prop diameter (mm)
Battery (mAh)
Electronics current (A) ≈0.15 no camera · ≈0.4 streaming
Overall efficiency η guess 0.18–0.30

–total weight

–max thrust (4 motors)

–thrust : weight

–hover throttle ≈

–hover current ≈

–flight time ≈ (80% of battery)

–payload left at 2:1

5.2 The four moves & the mixer

In plain words

Diagonal motors spin the same way: two clockwise, two counter-clockwise. Each spinning prop also twists the body the opposite way, like how your chair spins when you swing your arms. With the two directions balanced, the twists cancel. Then:

Throttle: all four faster means up.
Roll: left pair faster tips it right, and it slides right.
Pitch: back pair faster tips the nose down, and it moves forward.
Yaw: the clockwise pair faster makes the body twist counter-clockwise. Total lift stays the same, so it just turns on the spot.

The mixer is a tiny piece of code that turns "a bit more throttle, roll right a little" into four exact motor numbers.

Motor	Spin	Throttle	Roll right	Pitch up	Yaw right
M1 front-left	CW	+	+	+	−
M2 front-right	CCW	+	−	+	+
M3 rear-right	CW	+	−	−	−
M4 rear-left	CCW	+	+	−	+

"+" means the motor speeds up to make that move. Read a column: to roll right, the left motors (M1, M4) speed up. CW/CCW props are mirror-shaped, and one on the wrong motor blows air upward.

The maths

We use x forward, y right, z down. Positive roll is right-side-down, positive pitch is nose-up, and positive yaw is nose-right. Arm offset is a, and c is the torque-to-thrust ratio of the props:

Tτ_xτ_yτ_z=1111a−a−aaaa−a−a−cc−ccT₁T₂T₃T₄

The mixer is this matrix's inverse. The columns are orthogonal, so the inverse is easy. With t = T/4, r = τ_x/4a, p = τ_y/4a, y = τ_z/4c:

T1 = t + r + p − yT2 = t − r + p + yT3 = t − r − p − yT4 = t + r − p + y

When a motor would need more than 100% or less than 0%, you must squeeze the corrections. Keeping the drone level matters more than holding the exact throttle. You'll handle this "saturation" in v1.

5.3 Which way is up? (sensor fusion)

In plain words

The IMU chip has two sensors:

Gyroscope: measures how fast it's rotating. Smooth and instant, but tiny errors pile up, so it slowly drifts. It's like counting your steps with your eyes closed.
Accelerometer: feels gravity, so when still it knows exactly which way is down. But every vibration and every speed-up fools it. It's like a pendulum on a bumpy bus.

The trick: trust the gyro for quick changes, and let the accelerometer slowly pull out the drift. That blend is called sensor fusion.

The maths

Step 1, complementary filter (roll shown; pitch is the same). Integrate the gyro, then nudge toward the accelerometer's angle:

{\hat{φ}}_{k} = α ({\hat{φ}}_{k - 1} + ω_{x} Δ t) + (1 - α) φ_{acc}, α = \frac{τ}{τ + Δ t}

φ_{acc} = atan2 (a_{y}, a_{z}), θ_{acc} = atan2 (- a_{x}, \sqrt{a_{y}^{2} + a_{z}^{2}})

With Δt = 2 ms and α = 0.98, the time constant τ ≈ 0.1 s. Signs depend on how the chip is mounted, so you'll verify them by tilting by hand.

Step 2, quaternions. Angles break near 90° ("gimbal lock"). A quaternion q (4 numbers) represents any orientation and updates cheaply from the gyro rate ω:

\dot{q} = \frac{1}{2} q \otimes [0, ω + K_{p} e + K_{i} \int e]

That's the Mahony filter. e is the cross product between the gravity direction the accelerometer measures and the one the quaternion predicts. It's a PI controller that steers the estimate toward reality. Madgwick does the same job with gradient descent. Both are about 50 lines of code.

Step 3 (v3), the Extended Kalman Filter: the "proper" way to fuse many sensors with different noise and timing. See §5.6.

Watch outMotors vibrate at hundreds of hertz, which blinds the accelerometer. Use the IMU's built-in low-pass filter (on the MPU6050, the 42 Hz setting Gemini suggested is a fine start), soft-mount the IMU board if logs look noisy, and run the loop on a hardware timer so Δt is really 2 ms every time.

5.4 PID, the reflex

In plain words

Think of balancing a stick on your palm:

P (proportional): the further it leans, the harder you push back. Too much and you over-correct back and forth.
D (derivative): the brake. You see it swinging back fast and ease off before it overshoots. This kills the wobble.
I (integral): the patience. If it keeps leaning slightly one way (say the camera makes the nose heavy), I slowly builds up extra push until the lean is gone.

The maths

u (t) = K_{p} e + K_{i} \int e d t + K_{d} \frac{d e}{d t}, e = target - measured

In code, every Δt:

u_{k} = K_{p} e_{k} + K_{i} \sum_{j \leq k} e_{j} Δ t - K_{d} \frac{y_{k} - y_{k - 1}}{Δ t}

Three professional tricks:

Derivative on measurement (the −y term above, not +e) avoids a "kick" when you suddenly change the target.
Anti-windup: clamp the integral so it can't grow huge while the drone sits on the ground.
Low-pass filter the D term, because D amplifies noise.

PID sandbox: one tilting axis (like your string rig)

Each run: hold level, then the target jumps to 15° at 0.3 s and back to 0° at 2.0 s, then a finger tap hits it at 2.8 s. "Nose-heavy" is a constant lean force, like a camera on the front. The physics is a rough model of a v1-size drone. Units: Kp in % motor difference per degree.

Top: angle (solid) vs target (dashed). Bottom: motor command. The little drone replays the run.

–overshoot

–settle time (±1°)

–steady lean error

–tap recovery (±1°)

–motor jitter

Tuning recipe (rig first, never on the floor)

Ki = Kd = 0. Raise Kp until it starts to oscillate, then back off 30–50%.
Raise Kd until the oscillation dies and taps feel like hitting thick honey. Stop before motors sound rough.
Add a small Ki to remove leans.
Tune the inner rate loop first, then the angle loop (§5.5).

5.5 Loops inside loops (cascaded control)

In plain wordsA manager says "go to the door". A supervisor turns that into "walk forward at 1 m/s". A worker turns that into "lean forward 5°". The fastest reflex turns that into exact motor speeds. Each inner loop must be several times faster than the one outside it, or the chain wobbles.

Solid arrows = commands flowing toward the motors. Dashed = measurements feeding back. Only the slow outer loop may live on the laptop.

Golden ruleNever close a fast loop over Wi-Fi. Wi-Fi can pause for 50–100 ms at random. The rate, angle and altitude loops always run on the drone. The laptop only sends slow goals ("be at x = 1.2 m").

The maths: feed-forward

Controllers work better when you give them a good first guess. For altitude, start from the measured hover throttle and let the PID fix only the difference. A brushed motor's speed also drops as the battery sags, so scale throttle by battery voltage:

u_{throttle} \approx (u_{hover} + u_{PID}) \cdot \frac{V_{nominal}}{V_{battery}}

5.6 Knowing where it is

"Where am I?" gets answered in layers, from the cheapest and most local to the most global:

Question	Sensor	How it works (plain)	Good to know	Version
How high?	Laser time-of-flight (VL53L1X)	Fires invisible light at the floor and times the echo	Up to ~4 m in the dark, only ≈0.7–1.3 m in strong ambient light. Minimum 4 cm. Breakout ≈0.5 g.	v2
How high, outdoors?	Barometer (BMP390 / DPS310)	Air pressure drops as you climb	BMP390 relative accuracy ≈ ±25 cm. Prop wash fools it, so cover it with foam.	v4
Am I sliding?	Optical flow (PMW3901 / PAA3905)	A tiny camera like a computer mouse that watches the floor slide past	Works from ~8 cm up. Needs light and floor texture. Bitcraze's Flow deck (flow + laser) weighs 1.6 g.	v2
Where in the room?	Camera + printed markers (ArUco / AprilTag)	Knows each marker's size and position, so its shape in the image gives the drone's position	Runs on the laptop. Centimetre-level near markers. Arrives 0.1–0.3 s late.	v3
Where on Earth?	GPS (+ compass)	Times signals from satellites	A few metres accurate, outdoors only. Tiny modules: Flywoo GM10 Nano 2.2 g, HGLRC M100 Mini ≈2.7 g.	v4

The maths

Height with tilt correction: the laser points along the body, so when tilted it measures a slanted distance d:

h = d cos φ cos θ

Speed from optical flow: the floor's apparent angular speed times height gives ground speed. But tilting also makes the floor "move", so subtract the gyro's rotation:

v \approx h \cdot (Ω_{flow} - ω_{gyro})

Position from a marker (Perspective-n-Point): a 3D point X on a marker lands on pixel (u, v) through the camera matrix K (found by calibration) and the unknown pose (R, t):

suv1=K[ R | t ]X1p_camera = −R^⊤t

OpenCV's solvePnP finds R and t from the four corners of each marker.

Fusing everything: the Extended Kalman Filter. Keep a state x (position, velocity, tilt…) and its uncertainty P. Predict with the IMU every step, and correct whenever any sensor reports:

{\hat{x}}^{-} = f (\hat{x}, u), P^{-} = F P F^{⊤} + Q

K = P^{-} H^{⊤} {(H P^{-} H^{⊤} + R)}^{- 1}, \hat{x} = {\hat{x}}^{-} + K (z - h ({\hat{x}}^{-})), P = (I - K H) P^{-}

In plain words: K is a "trust dial". If a sensor is noisy (big R) it barely moves the estimate, and if you're unsure of the estimate (big P) the sensor gets a bigger say. Crazyflie's firmware has a well-documented EKF to study (§10).

Watch out: delaysMarker positions computed on the laptop arrive 100–300 ms late. Stamp every camera frame with the drone's clock and fuse the measurement at the time it was taken, not when it arrived. Skipping this is the #1 reason student "autonomous" drones oscillate.

5.7 Electronics & pins

The battery

In plain wordsA 1S LiPo is one lithium cell: 4.2 V when full, 3.7 V "nominal", empty around 3.3–3.4 V under load. It can dump huge current in a short time, which is great for motors and dangerous when shorted (§8). mAh is the size of the tank, and the C-rating is how fast it can safely empty: 250 mAh × 30C = 7.5 A on paper.

Fact	Number	Why it matters
Land when loaded voltage reaches	≈ 3.3–3.4 V	Lower than that hurts the cell, and the ESP32's 3.3 V supply starts to sag.
Store at	≈ 3.8 V	Full or empty LiPos age fast in storage.
Never below	3.0 V	Permanent damage. A puffed cell means retire it safely.
PH2.0 connector	≈ 4.5 A	Four motors can pull 4–6 A. Too small.
BT2.0 connector	≈ 9 A continuous	Use this. Batteries and charger must match.
250 / 300 mAh whoop pack	≈ 7.1 / 8.3 g	The heaviest single part of v1. Pick capacity with the calculator.

One motor channel (×4)

MOSFET (AO3400A): an electronic tap. A tiny 3.3 V signal on its gate lets several amps flow through the motor.
Schottky diode across the motor: a pressure-relief valve. When the tap shuts, the motor's coil tries to keep current flowing and spikes the voltage. The diode gives that current a safe loop. Without it you get resets or dead parts.
100 Ω gate resistor: softens the switching a little and stops ringing.
10 kΩ pull-down: keeps the motor OFF while the chip boots, when pins float. This is a real safety part.
PWM: flicking the tap on/off ~20,000 times a second (above human hearing). The motor feels the average, so 60% on-time gives roughly 60% power.

Layout rules that save you

Motor current goes battery → motor → MOSFET → ground on short, wide copper (≥ 1 mm traces or copper pours). It must never pass through the XIAO.
Ceramic capacitors (10–22 µF) near the XIAO's power pad and near the motors act as small local reservoirs for sudden current gulps.
IMU in the centre of the board, the Wi-Fi antenna away from motors and battery, and a solid ground plane underneath.

The four "languages" chips speak

Bus	Wires	Plain words	Used for
I²C	2 (SDA, SCL)	A party line: many chips share two wires, each answers to its own address	IMU (v1), height laser, barometer, compass, GPIO expander
SPI	3 shared + 1 "select" per chip	Faster. The brain points at one chip at a time with its select wire	Optical flow, microSD, IMU on custom boards
UART	2 (TX, RX)	A plain serial chat between exactly two devices	GPS, ExpressLRS receiver
ADC	1	Measures a voltage	Battery voltage (through a divider that halves it)

The pin wall

In plain wordsThe XIAO ESP32S3 has only 11 usable pins (D0–D10). With the camera, microSD and mic in use, just 8 are free. Only GPIO1–10 can measure voltage (ADC). The ESP32-S3 can route I²C, SPI and UART to almost any pin, so you get to shuffle. Pin planning is a real design puzzle, and it's the reason the Queen gets a custom chip-level board with 30+ free pins.

v1 XIAO, no camera

D0 GPIO1	Motor 1 FL
D1 GPIO2	Motor 2 FR
D2 GPIO3	Motor 3 RR
D3 GPIO4	Motor 4 RL
D4 GPIO5	I²C SDA (IMU)
D5 GPIO6	I²C SCL
D6 GPIO43	spare (UART TX)
D7 GPIO44	spare (UART RX)
D8 GPIO7	status LED
D9 GPIO8	spare
D10 GPIO9	battery (ADC)

v2 Sense: camera + SD

D0–D3	Motors 1–4
D4 GPIO5	battery (ADC)
D5 GPIO6	flow sensor select
D6 GPIO43	I²C SDA (IMU, laser, expander)
D7 GPIO44	I²C SCL
D8–D10	SPI: microSD and flow sensor share it
internal	camera, mic, SD select (GPIO21)
no pins left	payload magnet + LEDs go through an I²C GPIO expander

v4 XIAO, outdoor

D0–D3	Motors 1–4
D4/D5	I²C: IMU, barometer, compass
D6/D7	GPS (UART)
D8 GPIO7	ELRS receiver → drone
D9 GPIO8	drone → ELRS (telemetry)
D10 GPIO9	battery (ADC)
All 11 used. Video comes from an optional analog FPV camera, which needs power only.

The motor pins match the spin layout in §5.2. On v1, D8–D10 are free because the Sense board isn't attached. Fitting the Sense later turns them into SD-card lines, which is exactly Gemini's mistake.

5.8 Talking to it

Link	Talks to	Plain words	Range (rough)	Version
Wi-Fi (drone makes its own network)	Phone, laptop	Your phone joins the drone's Wi-Fi and opens a web page the drone serves	tens of metres	v1
ESP-NOW	Your own remote	ESP32-to-ESP32 messages with no router. Very low delay	similar to Wi-Fi	v3
ExpressLRS 2.4 GHz (CRSF)	RC radio	A hobby-grade long-range control link. A 0.45 g receiver sends 16 channels to the drone over UART	far beyond line of sight (stay legal, §8)	v4
Analog 5.8 GHz video	Goggles / Android USB receiver	Old-school TV-style video with almost no delay	hundreds of metres	v4

The phone app that isn't an appThe drone runs a tiny web server. Your phone joins its Wi-Fi network (say "PROJECT-BEE") and opens 192.168.4.1. The page shows two touch joysticks, an arm button, battery % and live video. Stick positions go out ~50 times a second over a WebSocket. It works on iPhone and Android with no app store. Outdoors (v4), you flip it around: the drone joins your phone's hotspot, so the phone keeps internet for map tiles.

Who's in charge? (non-negotiable safety)

Priority: RC radio / your remote > phone / laptop > autonomy. Touching the sticks always takes over.
Arming: motors can only start with throttle at zero plus a deliberate arm action. A power-up never spins props.
Link lost for 0.3 s → hold position or descend. Lost for 1 s → land. Tilt over 60° → motors off (it crashed). Low battery → land.
A kill switch on every controller.

5.9 Camera & video

In plain wordsThe camera chip compresses each picture into a JPEG by itself. The ESP32 just ships those pictures out over Wi-Fi (live view) or writes them to the microSD card (recording). Video is always a bit late (0.1–0.3 s), so it's for watching and AI, never for keeping the drone level.

Fact	Detail
Which camera arrives	Seeed now ships the Sense with an OV3660 or OV5640 (the OV2640 is discontinued). Espressif's camera driver supports all three.
What people have achieved	The ESP32-CAM_MJPEG2SD project: 25 fps from the camera and 20 fps to SD at VGA/SVGA, with streaming and recording together possible ("may stutter"). The hx-esp32-cam-fpv project on the same XIAO Sense + OV5640: 30 fps from 640×360 up to 1280×720 with 90–110 ms delay, recording 720p30 onboard.
Power	Seeed measured ≈ 155 mA average and ≈ 366 mA peak for its webcam demo.
Recording choices	On the drone's SD: best quality, independent of Wi-Fi. On the laptop: zero weight, but every Wi-Fi hiccup shows.
Analog FPV (v4)	An all-in-one camera + transmitter like the BetaFPV H02 (3.6 g, 25 mW, runs on 2.9–5.5 V, so straight off 1S). Watch it in goggles (near-zero delay) or on Android with a USB receiver (Eachine ROTG02: Android only, ≈ 100 ms).

The two-brain question (honest unknown)Can one ESP32-S3 fly and film? The flight loop is light, and Wi-Fi + camera live on the other core, so it should work. But Seeed's own ESP-FLY FAQ says its current firmware can't add the camera, so nobody has proven it on that design. You'll measure it on the bench (block 19) before trusting it. Plan B is two chips: a flight XIAO and a camera XIAO talking over two wires. It adds ≈ 3–4 g.

5.10 Firmware map: logic cards

You said you want to understand the code by its logic, not word by word. So every module gets a card: what it's for, what goes in and out, how often it runs, and how you prove it works. If you can explain every card to a friend without notes, you understand your drone.

Module	Its job (plain)	In → Out	How often	How you test it
Arming state machine	Decides whether motors may spin at all	buttons, link health, tilt, battery → DISARMED / ARMED / FLYING / FAILSAFE	every loop	Try to arm with throttle up. It must refuse.
Sensor drivers	Talk to each chip and turn raw bytes into real units	bus bytes → °/s, g, metres	IMU 500–1000 Hz, others 10–100 Hz	Board flat = 1 g down. Ruler = laser reading.
Estimator	Works out tilt, then height, speed, position	sensor values → attitude (quaternion), height, velocity	500 Hz	3D cube viewer, hand-carry tests
Commander	Picks who's in charge and produces one target	radio / phone / laptop / autonomy → setpoint	50–100 Hz	Switch a source off. The next one takes over, or failsafe.
Controllers	The reflexes: rate, angle, altitude, velocity, position	target + estimate → corrections	50–1000 Hz	Tilt test, tethered steps, logs
Mixer	Turns corrections into 4 motor numbers	throttle, roll, pitch, yaw → M1–M4	500 Hz	Check against the §5.2 table by hand
Motor output	Makes the PWM signals	0–1 → duty cycle	500 Hz	Spin test, props off
Comms	Radio in and out	packets ↔ commands, telemetry	50–100 Hz	Latency test, unplug test
Blackbox logger	Records everything for later plots	all variables → laptop / SD	50–500 Hz	Plot a flight in Python
Safety supervisor	Watches for trouble and triggers failsafes	tilt, link age, battery, sensor health → actions	100 Hz	Failsafe test matrix (block 27)
Camera task	Captures, streams and records	frames → Wi-Fi / SD	15–30 fps	fps + delay test (film a stopwatch)

Project folder layout

micro-drone/
  firmware/   flight code (PlatformIO)
  sim/        Python simulator
  ground/     laptop ground station
  web/        phone control page
  hardware/   KiCad boards
  cad/        3D-printed parts
  log/        build log + flight logs
  docs/       this plan, notes

Tools on your Mac

VS Code + PlatformIO, using the Arduino framework at first: friendly, and it still gives you FreeRTOS and ESP-IDF calls when you need them.
ESP-IDF, only to build the safety-net firmware (and maybe v5).
Python 3: numpy, matplotlib, pygame, opencv-contrib-python, later ultralytics + mediapipe.
KiCad 10 for PCBs. Onshape or Fusion for 3D parts.
Git + GitHub: commit every working step.

6The build plan

76 blocks of 3 days each. Every block has Learn (what to understand first), Do (what to build) and Done when (a test you can pass or fail, with no "kind of works"). Later versions are less detailed on purpose. We'll re-plan each one with what you learn on the way.

Your start date:

Loading plan…

Rules for every block

Write 5 lines in your build log at the end of each day: what worked, what broke, what you learned, numbers you measured, next step.
Commit to Git every time something works.
Props off for every bench test. Safety glasses for every powered test.
If a block takes longer, that's fine. Don't skip its "Done when" test.

7Shopping & budget

About these pricesEvery ₹ figure is a rough range for India in late 2026, not a quote. Prices and stock change weekly. Check 2–3 shops before buying. The only figures I checked directly are LionCircuits PCBs (₹799 for 5), XIAO Sense listings (≈ ₹1,300–2,200, patchy stock) and the radios' official US prices.

Order A: today (bench + v1)

Item	Why	Qty	≈ ₹
Tools & consumables (you keep these forever)
0.01 g digital scale (jewellery scale)	Weigh every part. Your most important tool.	1	350–600
No-clean flux (paste or gel)	Makes solder flow. SMD soldering is impossible without it.	1	200–400
Solder wire 0.5–0.6 mm, 63/37 leaded	Easiest for beginners (ventilate, wash hands)	1	200–400
Desoldering wick	Fixes mistakes	1	100–200
Fine tweezers (straight + curved)	Placing tiny parts	1 set	200–400
SMD soldering practice kit	20 practice joints before your real board	1	150–350
Kapton tape, IPA 99% + old toothbrush	Heat-proof masking; cleaning flux off	1 each	300–600
Safety glasses	Props and snipped leads go for your eyes	1	100–250
LiPo safety bag	Charge and store batteries inside it	1	300–600
1S LiPo charger, multi-port, BT2.0	Must match the battery connector	1	600–1,500
Breadboard, jumper wires, header pins	Bench prototyping	1 set	200–350
Silicone wire 26–28 AWG (red + black)	Battery leads, motor fixes	1 m each	100–250
Electronics
Seeed XIAO ESP32S3 Sense	The brain + camera (camera used from v2)	1	1,300–2,600
Seeed XIAO ESP32S3 (plain)	Flies v1 so the Sense stays safe. Later becomes your remote's brain.	1	800–1,200
MPU-6050 breakout (GY-521)	IMU for bench + v1. Matches the safety-net firmware.	2	150–300
AO3400A MOSFET (SOT-23)	Motor switches (+ spares)	20	100–200
Schottky diode, SOD-123 (e.g. 1N5819W / B5819W)	Flyback protection	20	100–200
0805 resistors: 100 Ω, 10 kΩ, 100 kΩ	Gate, pull-down, battery divider	20 each	100–300
Ceramic capacitors 0805/1206: 100 nF, 10 µF, 22 µF	Local power reservoirs	10 each	100–300
0716 coreless motors, 2 CW + 2 CCW	v1 motors (+ a full spare set)	2 sets	800–1,600
31 mm props matching the motor shaft (check 0.8 vs 1.0 mm)	They break, so buy plenty	3 sets	200–500
1S LiPo 250–300 mAh, ≥ 30C, BT2.0	~5 min each. 4 means continuous testing.	4	1,200–2,400
BT2.0 male pigtails	Battery lead on the drone	3	150–300
0805 LEDs	Status lights	10	50–100
Fabrication (in block 7)
v1 PCB × 5 (LionCircuits)	Your first board, ₹799 for 5, 5–6 days	5	800–1,000
3D-printing filament (PLA/PETG)	Thrust stand, motor clips, test frame	—	0–800
Order A total			≈ 8,650–17,700

Later orders

Order	When	What	≈ ₹
B · v2	block 17	VL53L1X laser (400–800), PMW3901 flow sensor (1,200–2,500), I²C GPIO expander (100–300), magnets / tiny electromagnet (100–300), v2 PCB ×5 (800–1,000), spare motors + props (500–1,000), 2 more LiPos (600–1,200). Only if thrust tests say so: 0720 motors + 40 mm props (600–1,200).	4,300–8,300
C · v3	block 29	Your remote: 2 thumb joysticks, 0.96" OLED, small LiPo, switch, perfboard (uses the plain XIAO). Printed markers + checkerboard.	550–1,200
D · v4	block 39	ExpressLRS radio: BetaFPV LiteRadio 3 (US$32.99 official) or Radiomaster Pocket (US$71.50) (3,500–8,000). ELRS 2.4 GHz receiver (1,000–1,800). Mini GPS + compass (1,800–3,000). Barometer breakout (400–900). 8520 motors ×2 sets + 55–65 mm props (800–1,600). 1S 450–600 mAh LiPos ×3 (900–1,800). v4 PCB ×5 (800–1,000).	9,200–18,100
D+ optional	block 50	Analog FPV: AIO camera like the BetaFPV H02 + Android USB receiver (ROTG02) or cheap goggles	4,000–6,500
E · v5	block 52	2–3 candidate tiny motor/prop sets (1,000–2,000). 1S 100–150 mAh LiPos ×4 (800–1,600). Factory-assembled Queen boards ×5 incl. shipping and duty (3,000–6,000). New backpack sensors if not reusing v2's (0–2,500).	4,800–12,100
F · v6	block 68	Motors, props and batteries for Queens #2 and #3 (their boards come from Order E's batch). Dock: contacts + small charger module.	2,300–5,000

Budget summary

Path	Through v3	Through v4	All the way (v6)
Cheapest (low end of every range, no FPV)	≈ ₹13.5k	≈ ₹22.7k	≈ ₹30k
Comfortable (mid-range parts, spares, FPV)	≈ ₹20k	≈ ₹34–39k	≈ ₹46–52k

Straight answerYour ₹15–35k comfortably covers v0–v4. Getting all the way to v6 inside ₹35k is possible only on the cheapest path. Decide at block 52 whether to stretch the budget or trim. Ways to trim:

Skip analog FPV (saves ₹4–6.5k).
Use the LiteRadio 3, not the Pocket, or fly v4 with your own remote and skip the RC radio (saves ₹4.5–10k).
Reuse v2's laser + flow sensor on the Queen's backpack.
Buy the 5 Queen boards in one factory order, which gives you the swarm boards for free.
Bundle orders so you pass free-shipping limits.

Where to buy

Inside India

General electronics: Robu.in, Evelta, Thingbits, Probots, Robocraze, Quartz Components
Whoop / FPV parts (tiny motors, 1S packs, chargers, ELRS): whoopmastersindia.com, uavstore.in, quadkart.in, fioparts.com, thefpvproject.com. I haven't checked their stock.
XIAO ESP32S3 Sense: stock is patchy. Check Robocraze, ElectronicsComp, Hubtronics and element14 India.
PCBs: LionCircuits (₹799 for 5, 5–6 days, also does assembly), PCB Power (fab + assembly)

International (for chips & v5)

DigiKey India: free delivery from ₹7,000, with duty and GST paid on delivery (≈ 4 days)
Mouser India, element14 India, LCSC (LCSC needs ID/KYC for DHL)
JLCPCB: cheap factory assembly. India has no duty-free limit, so you pay import duty plus KYC.
AliExpress doesn't sell to India.

Don't import a whole droneSince February 2022, India bans importing drones, including as kits (complete, semi-knocked-down or completely-knocked-down). Components are free to import. My reading is that a ready-to-fly whoop or a full kit like ESP-FLY counts as a drone, so buy parts, not kits. You can still study ESP-FLY's open-source files.

8Safety & the law

Safety rules (from day 1)

LiPo batteries

Charge inside the LiPo bag, on a non-flammable surface, never unattended, never overnight.
Never charge or fly a puffed, dented or hot cell.
Battery leads must never touch: a short means instant heat and possibly fire.
Unplug the battery before soldering anything.
Store at ≈ 3.8 V. Keep a bucket of sand nearby.

Props, soldering, power-up

Props off for all bench and code tests. Safety glasses whenever props are on.
First flights tethered, then over a bed or foam mat, away from faces and pets.
Ventilate flux fumes, wash hands after leaded solder, and keep the iron in its stand.
Before the first power-up: multimeter check that battery+ isn't shorted to ground. First power from USB, with no motors attached.

The law in India (what the rule text says)

I'm not a lawyer. This is what the published rule texts say as of October 2026 (sources in §10). Check DGCA's Digital Sky platform yourself before any outdoor flight.

Topic	What the rules say	What it means for you
Category	Drone Rules 2021: nano = up to 250 g	Every version you'll build is nano (Forager ≈ 50 g).
Pilot certificate, type certificate, insurance	Not required for nano	Good news.
Registration (UIN)	Rule 14: no flying without Digital Sky registration + UIN "unless exempted", and the text has no nano exemption. The popular "nano needs no UIN, fly up to 50 ft" claim comes from the old 2018 rules, which were repealed.	Unclear for home-built nanos in practice. Sort this out before v4's outdoor flights.
Where outdoors	Green zone: up to 120 m (400 ft) with no permission (60 m in the 8–12 km band around airports). Yellow: ATC permission. Red: central government only. Check the Digital Sky map before every flight. Fines up to ₹1 lakh.	Your v4 geofence caps height well below the legal limit.
R&D exemption	Rule 42: recognised educational institutions / R&D bodies can fly on their own premises in green zones without UIN or permissions.	If you're still enrolled, ask your department whether v4 can fly on campus under the college's R&D umbrella.
Coming changes	The draft Civil Drone Bill 2025 would require registration for every drone. It appears to still be a draft.	Re-check when v4 starts.
Indoors	The rules are written for outdoor airspace.	Your own room or hall is the practical safe zone for v1–v3 and v5–v6.
Radio	2.4 GHz is licence-exempt. 5.725–5.875 GHz is licence-exempt. The 865–868 MHz rules need ≥ 58 hop channels and a low duty cycle.	Use 2.4 GHz ExpressLRS, not the 866 MHz version (it doesn't meet those rules). For FPV, only Raceband channels R3–R6 (5732–5843 MHz) are inside the exempt band.
Imports	Complete drones and kits are banned. Components are free.	Buy parts (§7).

9Learning with AI, without depending on it

In plain wordsI'll write a lot of the code with you. That's fine, as long as you own the logic. The test is simple: could you explain any module to a friend, and predict what happens when you change a number? If yes, it's yours.

The loop for every module

Spec: you write 3–5 lines in plain words: what it does, what goes in and out, how you'll test it.
Predict: roughly what steps should the code take?
Build: I write it in small, commented functions.
Explain back: you walk me through it as its logic card (§5.10), and I point out gaps.
Change one thing yourself: a gain, a threshold, a rate. Predict the effect, then run it.
Test: pass the block's "Done when".
Log: 5 lines and a Git commit.

Rules

Never keep code you can't explain at logic-card level.
When debugging, write your guess first, then ask.
Ask "why this number?" about every magic number.
Once per version, rebuild one small module with no AI (for example the complementary filter) as a self-test.

Things to say to me

"Here's my spec for ___. Build it in small functions, then quiz me."
"Explain this file as a logic card."
"I think the bug is ___ because ___. What test would prove it?"
"Review my schematic PDF against the v1 checklist."
"Give me 5 questions to check I really understand ___."

10Resources

Links checked on 24 Oct 2026. The block numbers say when each one becomes useful.

Flight control from scratch

Carbon Aeronautics manual + videos: a quad built from zero, explained step by step (blocks 2–16)
dRehmFlight: very readable flight-controller code (blocks 8–13)
Joop Brokking YMFC-AL: the classic Arduino quad series
MATLAB Tech Talks: Drone Simulation and Control (5 short videos, blocks 2 and 9)
Bitcraze docs: controllers · state estimators
ESP-Drone (docs) and Seeed ESP-FLY: your safety net and reference design

Maths, estimation, simulation

Kalman and Bayesian Filters in Python (Roger Labbe): free interactive book (block 31)
Madgwick's report + the Fusion library (block 4)
Beard & McLain, Small Unmanned Aircraft: free draft book + simulators
gym-pybullet-drones: Crazyflie simulation, and RL later

11Glossary

1S: One lithium cell in the battery (3.7 V nominal). 2S = two in series, and so on.
ADC: Analog-to-digital converter: lets the chip measure a voltage.
Arming: The deliberate act of allowing the motors to spin. Disarmed = motors can't run.
ArUco / AprilTag: Printed black-and-white square markers a camera can find and measure precisely.
Blackbox: A flight recorder: logs every sensor and controller value so you can plot what happened.
Brushed / brushless: Brushed (coreless) motors are simple and run straight from a MOSFET. Brushless are stronger but need an ESC.
BT2.0 / PH2.0: Tiny battery connectors. BT2.0 handles much more current.
C-rating: How fast a battery can safely discharge, as a multiple of its capacity.
Cascade: Control loops stacked inside each other: position → velocity → angle → rate.
Complementary filter: Blends gyro (fast, drifts) with accelerometer (slow, noisy) to estimate tilt.
Coreless motor: A tiny brushed motor with no iron core. Light and fast to respond. "0716" = 7 mm × 16 mm.
CRSF: The serial protocol ExpressLRS receivers use to send stick values to the flight controller.
CW / CCW: Clockwise / counter-clockwise (seen from above). Props come in mirrored pairs.
DLPF: Digital low-pass filter: smooths out fast vibration from a sensor's readings.
EKF: Extended Kalman filter: the standard way to fuse many sensors with different noise and timing.
ELRS (ExpressLRS): Open-source long-range radio control system used by most hobbyists.
ESC: Electronic speed controller: drives brushless motors.
ESP-NOW: Espressif's direct chip-to-chip radio messages: no router, low delay.
ESP32-S3: The microcontroller brain: two cores, Wi-Fi + Bluetooth, a camera interface.
Failsafe: What the drone does automatically when something goes wrong (lost link, low battery).
Feed-forward: Giving a controller a good first guess (like hover throttle) so it only corrects the difference.
Firmware: The program that runs on the drone's chip.
Flyback diode: Gives a motor coil's current a safe path when the switch turns off.
FPV: First-person view: flying by watching the drone's live camera.
FreeRTOS: The small operating system inside the ESP32 that runs several tasks at once.
Gerber: The files a PCB factory needs to make your board.
GPIO: A general-purpose pin the chip can read or drive.
Gyroscope: Measures rotation speed (°/s).
I²C, SPI, UART: Ways chips talk to each other (§5.7).
IMU: Inertial measurement unit: a gyro + accelerometer chip.
KiCad: Free software to design circuit boards.
LiPo: Lithium-polymer battery: light and powerful, and must be treated with care.
mAh: Battery capacity. 250 mAh ≈ 0.25 A for one hour.
Mixer: Turns throttle / roll / pitch / yaw into four motor commands.
MOSFET: An electronic switch. A small voltage on its gate controls a large current.
Optical flow: Measuring motion by watching how the image of the floor slides.
PCB / PCBA: Printed circuit board / one that the factory has already assembled with parts.
PID: Proportional-integral-derivative controller: the reflex (§5.4).
PnP: Perspective-n-point: finding the camera's pose from known points in an image.
PSRAM: Extra memory on the ESP32 module, used for camera frames.
PWM: Pulse-width modulation: fast on/off switching whose average sets motor power.
Quaternion: Four numbers that represent orientation without the 90° glitch that angles have.
Rate / angle mode: Angle mode self-levels (easy). Rate ("acro") mode holds rotation speed (for flips).
Schottky diode: A fast diode with a low voltage drop.
Setpoint: The target value a controller tries to reach.
SLAM: Simultaneous localisation and mapping: building a map while working out where you are in it.
SMD: Surface-mount device: tiny parts soldered flat onto pads.
Telemetry: Data the drone sends back: battery, attitude, position.
Thrust-to-weight: Max thrust ÷ weight. Aim for 2:1 or more.
ToF: Time-of-flight laser distance sensor.
UIN: Unique identification number for a registered drone in India.
VTX: Video transmitter (analog FPV).
Wheelbase: Diagonal distance between motor centres. "65 mm whoop" refers to this.
Whoop: A tiny ducted indoor drone class.

Made for your Project Bee, 24 Oct 2026. When reality disagrees with this plan, reality wins. Tell me and we'll update it.

0Start here

1What we agreed

2Gemini's plan, reviewed

3The big picture

The version ladder

Why this order?

What's inside (v2 shown)

Your timeline

4Everything it could do

Flying

Seeing

Knowing where it is

Doing things by itself

Talking to it

Carrying things

Teams

Maths & Computing specials

5How it works inside

5.1 Why it flies, and why tiny is hard

The 4 cm box rule

Thrust-to-weight and hover throttle

Weight, thrust & flight-time calculator

5.2 The four moves & the mixer

5.3 Which way is up? (sensor fusion)

5.4 PID, the reflex

PID sandbox: one tilting axis (like your string rig)

5.5 Loops inside loops (cascaded control)

5.6 Knowing where it is

5.7 Electronics & pins

The battery

One motor channel (×4)

The four "languages" chips speak

The pin wall

v1 XIAO, no camera

v2 Sense: camera + SD

v4 XIAO, outdoor

5.8 Talking to it

5.9 Camera & video

5.10 Firmware map: logic cards

Project folder layout

Tools on your Mac

6The build plan

7Shopping & budget

Order A: today (bench + v1)

Later orders

Budget summary

Where to buy

Inside India

International (for chips & v5)

8Safety & the law

Safety rules (from day 1)

LiPo batteries

Props, soldering, power-up

The law in India (what the rule text says)

9Learning with AI, without depending on it

The loop for every module

Rules

Things to say to me

10Resources

Flight control from scratch

Maths, estimation, simulation

Electronics & PCB

ESP32-S3, camera, radio

Coding basics

Datasheets & law

11Glossary