Quadrotor Control Guide

This guide covers quadrotor flight control from the rotors up. It does three things:

It characterizes six performance strata, from 15 km/h indoor research vehicles to 720 km/h speed-record airframes. For each it gives the size, weight, motors, and thrust, and explains why those parameters follow from the speed.
It derives what changes for the controller between strata: which physical effects dominate, how fast the loops must run, and which model terms the controller and the simulator need.
It lists the control problems reflex must solve, with a design question, an implementation home, and a test plan for each, and maps every problem onto the strata where it matters.

It sits above the component pages (PID, quadrotor model) and links the roadmap issues that carry the work.

Conventions

The conventions are those of the quadrotor model:

Frames: the world frame has z up. The body frame is x forward, y left, z up (FLU).
Rotor order: PX4 Quad X. Rotor 0 is front-right CCW, rotor 1 rear-left CCW, rotor 2 front-left CW, rotor 3 rear-right CW.
Symbols:
- $n = T_{max}/(m g)$ is the thrust-to-weight ratio.
- $q = \tfrac12 \rho v^2$ is the dynamic pressure, with $\rho = 1.225\ \mathrm{kg/m^3}$ at sea level.
- $C_D A$ is the drag area.
- $J = V/(n_{rev} D)$ is the propeller advance ratio: airspeed $V$ along the rotor axis, rotor speed $n_{rev}$ in rev/s, diameter $D$ .

The control stack in one page

 planner / MPC (companion, 100–200 Hz)
        │  collective thrust + body rates (CTBR), optional feed-forward torques
        ▼
 attitude loop  (FC, 250–1000 Hz)      quaternion error → rate setpoint
        ▼
 rate loop      (FC, 1–8 kHz)          PID or INDI on filtered gyro → torque demand
        ▼
 mixer          (FC, same rate)        wrench → four rotor thrusts, saturation priority
        ▼
 motor command  (ESC, DShot)           thrust → rotor speed / throttle, battery compensation
        ▼
 rotors + airframe  ← modelled by QuadrotorModel in SITL

reflex owns everything from the attitude loop down, plus the vehicle model. The planner and the state estimate come from cortex and racing-drone. Across the six strata the stack keeps the same shape. What changes is the physics inside each box, the rates the loops must run at, and which saturations bind.

Performance strata

A stratum is named by a nominal speed near the top of its vehicle class’s operating envelope: the speed the class races or sprints at in normal use, not a one-off record run. Cruise speeds are lower; a camera drone, for example, is rated for endurance well below its top speed [4]. Each stratum has reference vehicles. Manufacturer and record figures are cited. Values marked ≈ are this guide’s engineering estimates, computed from the cited figures with the relations in why the parameters are what they are.

Summary

Stratum	Speed	Reference vehicle	All-up mass	Size	Props	Motors	Power system	Thrust / weight
S0 ultra-slow	15 km/h (4 m/s)	Crazyflie 2.1 / 2.1 Brushless (research nano)	27 g [1]; 34 g brushless [2]	92 mm motor-to-motor [1]	45–55 mm, 2-blade	7×16 mm coreless brushed [1]; 08028 10000 KV brushless [2]	1S LiPo, 7–10 min [1, 2]	≈ 2 brushed; ≈ 3.5 brushless (4 × 30 g [2] on 34 g)
S1 slow-slow	30 km/h (8 m/s)	65 mm brushless tiny whoop (BetaFPV Meteor65)	≈ 31 g: 23 g dry [3] + 1S 300 mAh pack	65 mm motor diagonal, ducted	31–35 mm, 3-blade [3]	0802, 19 500–22 000 KV, 1.8 g each [3]	1S 300 mAh, ≈ 4.5 min [3]	≈ 3–4
S2 slow	60 km/h (17 m/s)	Consumer camera drone (DJI Mini 4 Pro: 57.6 km/h max [4])	< 249 g [4]	≈ 250 mm diagonal, folding	≈ 6”, 2-blade, low pitch	small outrunners, efficiency-wound	2S Li-ion, ≈ 30 min	≈ 2
S3 medium	120 km/h (33 m/s)	5”–6” racing quad (MonoRace A2RL 2025 drone; DCL racer)	966 g [5]; DCL rule 850–1000 g [6]	≥ 360 mm motor diagonal (DCL) [6]	5.1”, 3-blade [5]	2207-class, 1750–1950 KV, ≈ 1.7–1.9 kg static thrust each [7]	6S LiPo, ≈ 2 min	≈ 7 (MonoRace manoeuvres reach ≈ 7 g [5])
S4 fast	360 km/h (100 m/s)	Red Bull camera drone (filmed F1 at speed)	< 1 kg [8]	streamlined fuselage	high pitch; rotors up to 42 000 rpm [8]	high-KV racing motors, partly unreleased [8]	LiPo, ≈ 3 min [8]	≈ 10+ (100 → 300 km/h in ≈ 2–4 s [8])
S5 ultra (torpedo)	720 km/h (200 m/s)	Speed-record quads: Peregreen 2 / 3 / V4	2.77 kg (Peregreen 3) [9]	single-piece streamlined body, canted rotors	6” (V4) [10], very high pitch	T-Motor 3115 900 KV (P2) [11]; 3120 900 KV (V4) [10]	12S LiPo, 120 A ESCs (P2) [11]; ≈ 16 kW (P3) [9]; < 2 min	≈ 10–15

720 km/h is beyond the current record. Peregreen V4 holds the Guinness quadcopter record at 657.59 km/h, set on 11 December 2025 [10, 12]. Peregreen 3 flew 585 km/h on about 16 kW [9]. S5 is therefore a projection: holding the Peregreen 3 drag area, 200 m/s needs about 30 kW of electrical power (power ceiling). The speed column for S5 is a design target, not an observed operating point.

S0 — ultra-slow, 15 km/h

Use. Indoor research, swarms, flight near people, and learning-based control experiments.

Why it looks like this.

The vehicle is light because low mass keeps collision energy small and crashes cheap, not because of speed. Light is not harmless: spinning props still cut, so flight near people still needs prop guards and separation.
Its speed is set by the mission and the firmware, not by physics. A 3.5 thrust-to-weight brushless Crazyflie could fly much faster. Indoors, 15 km/h already covers a room in two seconds, and the motion-capture or flow-deck estimate degrades with speed.

Control character.

The dynamics are hover-dominated, so small-angle linearization works. Dynamic pressure is about 11 Pa, so aerodynamic drag is a small disturbance.
Rotor dynamics are slow relative to the airframe. Coreless brushed motors have a time constant of about 65–72 ms [13], against 20–25 ms for racing motors [5]. The rate loop must therefore respect motor lag explicitly.
The dominant disturbances are ground effect, ceiling effect, and downwash between vehicles in a swarm.
Thrust is very sensitive to battery voltage on 1S, so voltage compensation is essential [14].

S1 — slow-slow, 30 km/h

Use. Indoor FPV, proximity inspection, and confined-space flight.

Why it looks like this.

Ducts protect the props and add a little static thrust, at the cost of drag.
Very high KV on 1S spins tiny props fast enough for a thrust-to-weight ratio of 3–4.

Control character.

The airframe is the most agile of all strata in angular acceleration, because angular acceleration scales as $1/L^2$ (scaling). The rate loop needs high bandwidth.
Gyro noise and frame resonances sit close to the control band, so filtering (reflex#21) is critical.
Wall and duct aerodynamics matter.

S2 — slow, 60 km/h

Use. Aerial photography, mapping, delivery, and inspection.

Why it looks like this.

The vehicle is optimized for endurance, not agility. It uses large, low-pitch, low-disc-loading props, low-KV efficient motors, Li-ion cells, and a thrust-to-weight ratio near 2.
The class tops out near the stratum speed (16 m/s, 57.6 km/h, in sport mode [4]), enough to fight a 10 m/s wind with margin. It cruises far slower: DJI rates the maximum flight time at 21.6 km/h [4].
Regulation drives the mass: 249 g is the threshold for lighter rules in many jurisdictions [4].

Control character.

The controller is tuned for smoothness and disturbance rejection: gimbal stability, wind gusts, and position hold on GNSS.
Attitude excursions are small (< 35°).
The binding constraint is thrust margin. At $n \approx 2$ a gust plus a climb demand can saturate the motors, so the mixer’s thrust priority matters (reflex#19).

S3 — medium, 120 km/h

Use. Drone racing, human and autonomous. This is the stratum racing-drone targets for A2RL.

Why it looks like this.

At $n \approx 7$ the vehicle accelerates at up to 7 g and can hold about 80° of tilt at full thrust. That is what gate-to-gate racing needs, with turns of 20–30 m radius at 30 m/s.
The 5.1” props and 2207 motors on 6S are a mature sweet spot. Static thrust is high, the props survive crashes, and the pitch speed (about 60 m/s for a 4.1” pitch at about 34 000 rpm loaded, ≈) sits above the operating speed with useful margin.
Mass is held near 1 kg by league rules [6] and carries the compute (a Jetson-class companion on A2RL drones).

Control character.

This is the first stratum where aerodynamics shape the trajectory:
- Dynamic pressure reaches about 680 Pa.
- Rotor drag, which is proportional to velocity, and body drag change the achievable acceleration. MonoRace identifies linear and quadratic drag terms for its EKF and policy [5].
- Thrust lapses with advance ratio: the same command gives less thrust at 30 m/s than in hover.
Saturation is the normal operating condition, not an exception. Time-optimal flight rides the thrust limit, so mixer priority and integrator anti-windup decide whether the vehicle keeps attitude authority.
IMU saturation. MonoRace hit the 16 g accelerometer limit and had to fall back to the model [5].
Latency budget. 10 ms is 0.33 m of travel. The racing-drone end-to-end budget is ≤ 29 ms.

S4 — fast, 360 km/h

Use. Chase filming of race cars and motorcycles, and high-speed interception.

Why it looks like this.

To reach 100 m/s the airframe stops being a frame with props and becomes a fuselage. Dynamic pressure is about 6 kPa, nine times S3, so every square centimetre of frontal area counts.
At 42 000 rpm (700 rev/s) [8] the propeller needs a pitch of at least 5.5” just to reach the zero-thrust advance ratio at 100 m/s. To still produce thrust there it needs about 7” of pitch or more, which means a pitch-to-diameter ratio well above 1, unlike the 0.8 of a racing prop (≈).
High-pitch props give poor static thrust per watt, so the vehicle is a compromise between launch acceleration and top speed.

Control character.

The vehicle flies at high tilt (≈ 70–85°) for long periods, and its envelope reaches toward 90°, so attitude representation is no longer optional. Euler angles are singular at 90° pitch, and the controller must work on $SO(3)$ or with quaternions.
The body produces aerodynamic lift and moments that the rotors must trim. Gains that suit hover are wrong at 6 kPa, so gain scheduling with dynamic pressure (or with airspeed) is needed.
Turn radius at 5 g is about 200 m: the vehicle flies lines and arcs, not gates.
Estimation is GNSS-aided at high speed, and 10 ms of latency is 1 m of travel.

S5 — ultra / torpedo, 720 km/h

Use. Speed records and research into the limits of electric multirotor flight.

Why it looks like this. Every design choice serves top speed in a straight line:

The rotors are canted forward relative to a torpedo-shaped body, so the body is aligned with the airflow at the cruise tilt. Here $\theta$ $θ$ is the tilt of the rotor thrust from vertical; because of the cant, the body attitude differs from it.
- At the drag–thrust limit, full thrust gives $\theta_{max} = \arccos(1/n)$ , roughly 84° for $n = 10$ .
- That is a limit, not the cruise point. With the drag area estimated in the power ceiling and Peregreen 3’s 2.77 kg, drag at 200 m/s is about 76 N against 27 N of weight, so the rotor-thrust tilt is $\arctan(D/mg) \approx 70°$ before body lift (≈). S5 is limited by power and tip Mach, not by the drag–thrust balance.
Props are small (6” on V4 [10]) with very high pitch, on 900 KV motors at high voltage.
Batteries are chosen for power density, not energy. Peregreen 3 draws about 16 kW for under two minutes and needed water cooling after a prototype caught fire [9].
Mass goes up, to 2.77 kg on Peregreen 3: the extra inertia improves stability at speed [9], and the power system is heavy.

Control character.

Transonic propeller tips. A 6” prop at 0.85 × the 12S no-load speed of a 900 KV motor (≈ 640 rev/s) has a rotational tip speed of about 310 m/s. Combined with 200 m/s of forward flight, the helical tip speed is about 370 m/s, Mach 1.08 (≈).
- Thrust and torque coefficients change abruptly near Mach 1, so $T = k_f \omega^2$ no longer holds even approximately.
- Holding the helical tip below Mach 0.9 at 200 m/s caps the rotational tip speed near 235 m/s. Reaching 240 m/s of pitch speed then needs a pitch-to-diameter ratio above 3 (≈).
- This is the physical wall that makes 720 km/h harder than the current record. The Peregreen 3 builders cite the same constraint: they chose custom extreme-pitch APC props to keep the blade tips from going supersonic [16].
The thrust map is a function of airspeed. A $C_T(J)$ thrust model is mandatory in both the controller and the simulator.
Control authority shrinks at speed. The motors are near their power limit, so the differential thrust left for attitude is small. Mixer saturation priority is a flight-critical feature.
Structure and thermal. Aeroelastic flutter, ESC and motor thermal limits, and battery sag over a two-minute run are part of the control envelope.
Very large turn radius. At 5 g the radius is about 820 m, and 10 ms of latency is 2 m of travel.

Why the parameters are what they are

Four ceilings bound a quadrotor’s speed. Each stratum is limited by a different one, and the vehicle parameters follow from which ceiling binds.

1. The drag–thrust balance

In steady level flight at tilt $\theta$ , thrust balances weight vertically and drag horizontally:

T\cos\theta = m g, \qquad T\sin\theta = \tfrac12 \rho\, C_D A\, v^2 \;\;\Rightarrow\;\; v_{max} = \sqrt{\frac{2\, m g \sqrt{n^2 - 1}}{\rho\, C_D A}} .

Speed grows only as $\sqrt{n}$ and as $1/\sqrt{C_D A}$ . Doubling the speed needs four times $n$ , or a quarter of the drag area. This is why S4 and S5 spend their design effort on the fuselage. For the small strata this ceiling lies far above the operating speed, which the mission sets instead.

2. The propeller pitch-speed ceiling

A fixed-pitch propeller’s thrust falls as the axial inflow approaches its pitch speed $V_p \approx P \cdot n_{rev}$ , and reaches zero near the zero-thrust advance ratio $J_0$ (a little below $P/D$ for small props). To fly at speed $v$ :

P \gtrsim \frac{v}{n_{rev}} \quad\text{(plus slip margin)}.

S3: about 60 m/s of pitch speed limits a racing quad to the 140–160 km/h region quoted for racers [6, 15]. That ceiling, not the body drag, is what binds.
S4 and S5: reaching these speeds needs high pitch-to-diameter ratios, and the price is poor static thrust per watt.

3. The power ceiling

The electrical power to overcome body drag is $P_{el} = \tfrac12 \rho\, C_D A\, v^3 / \eta$ , where $\eta = \eta_{ESC}\,\eta_{motor}\,\eta_{prop}$ is the overall electrical-to-thrust-power efficiency.

A rough drag-equivalent area. Peregreen 3’s reported figures are a 16.2 kW peak drive-system draw and a 585 km/h record speed [9, 16]. They are not stated to come from the same steady, level-flight condition, and total electrical power also pays for lift and induced losses, not body drag alone. Treating the peak draw as the power at 162.5 m/s, with $\eta \approx 0.5$ (about 0.85 for ESC and motor times 0.6 for the propeller), gives a drag-equivalent area $C_D A \approx 0.0031\ \mathrm{m^2}$ : an upper-bound-style lumped figure, about a 9 × 9 cm frontal area at $C_D \approx 0.37$ (≈). For comparison, Peregreen 2 drew 8.9 kW at 1.85 kg for 480 km/h [16].
At 200 m/s, extrapolating with that lumped area and the same efficiency gives about 30 kW of electrical power: 1.86 times the power for 1.23 times the speed (≈). The $v^3$ ratio holds for any constant $\eta$ ; the absolute figures inherit every assumption above.
Consequences. The battery and ESCs are sized for peak power, the vehicle is heavier, and flight time is minutes.

4. Scaling with size

Scale the airframe geometrically by a length $L$ and hold the propeller tip speed (a Mach limit), so $n_{rev} \propto 1/D$ :

Quantity	Scaling	Consequence
Mass $m$	$L^3$	—
Thrust $\propto \rho n_{rev}^2 D^4$	$L^2$	$n \propto 1/L$ : small vehicles have spare thrust
Inertia $J$	$L^5$	—
Torque $\propto T L$	$L^3$	angular acceleration $\propto 1/L^2$ : small vehicles flip fastest
Natural frequencies	$\propto 1/L$	small vehicles need faster loops and better filtering

This is why S0 and S1 are the most agile in rotation yet the slowest in translation. Their speed is set by use and by drag on a bluff body, not by thrust. It is also why S5 grows heavier: the inertia damps disturbances at speed, and the motors that deliver 16 kW have mass of their own.

Per-stratum physics at a glance

	S0	S1	S2	S3	S4	S5
Speed [m/s]	4.2	8.3	16.7	33.3	100	200
Dynamic pressure $q$ [Pa]	11	43	170	680	6 100	24 500
Travel in 10 ms	4 cm	8 cm	17 cm	33 cm	1.0 m	2.0 m
Turn radius at 5 g	0.4 m	1.4 m	5.7 m	23 m	200 m	820 m
Typical rotor-thrust tilt	< 20°	< 45°	< 35°	up to 80°	≈ 70–85°	≈ 70° at 200 m/s, before body lift
Binding ceiling	mission	mission / drag	mission / endurance	pitch speed, thrust	drag, pitch	power, tip Mach
Is $T = k_f\omega^2$ adequate?	yes	yes	yes, with care	no: rotor drag, $C_T(J)$	no	no: transonic

What the current reflex model covers

QuadrotorModel implements:

rigid-body 6-DOF dynamics;
first-order motor lag;
the static thrust law $T = k_f\omega^2$ and the yaw-torque law $k_m T$ ;
linear and quadratic body drag;
battery sag;
ground contact.

That set is adequate for S0–S2 in free flight away from surfaces, and for S3 at low speed. It has ground contact but no ground or ceiling effect (E4), which is critical for S0 and S1, so near-surface scenarios in those strata are model-limited until E4 lands. Above about 15 m/s the effects listed under aerodynamic model terms become first-order, and SITL results there must be read as optimistic until they are modelled. Each term is a separate model feature with its own verification and validation pass.

Control problems to design, implement, and test

Each problem below lists:

Design: the question that must be answered before code. Stateful types are specified model, invariants, and contracts first.
Implement: where the code lives.
Test: the evidence that closes it.
Strata: where it matters (● critical, ○ relevant, – negligible).

A. Allocation and actuation

A1. Mixer and saturation priority — reflex#19.

Design:
- Closed-form inverse of the X allocation matrix.
- Priority order: roll/pitch, then collective thrust, then yaw.
- Two-sided thrust shift (airmode).
- Saturation report for anti-windup.
Implement: control/mixer.hpp.
Test:
- Round trip inside the limits.
- Corner cases of the thrust envelope.
- Cross-check of the torque signs against QuadrotorModel.
Strata: S0 ○ · S1 ○ · S2 ● · S3 ● · S4 ● · S5 ●

A2. Thrust linearization and battery compensation.

Design: command → thrust map, with thrust ∝ throttle² and a voltage-sag correction, as in Bitcraze’s thrust compensation [14].
Implement: a motor-output stage after the mixer.
Test: SITL with battery sag enabled. The hover throttle must stay constant over a full pack within ±3 %.
Strata: S0 ● · S1 ● · S2 ○ · S3 ● · S4 ● · S5 ●

A3. Motor dynamics compensation.

Design: feed-forward lead $1 + \tau s$ on the thrust command, or INDI with rotor-speed feedback (DShot telemetry).
Implement: motor-output stage, or the INDI rate loop (B3).
Test:
- SITL step response with $\tau$ in the range 20–70 ms.
- Phase margin measured, not assumed.
Strata: S0 ● · S1 ○ · S2 ○ · S3 ● · S4 ● · S5 ●

A4. Motor-loss detection and three-rotor fallback (later phase).

Design: detect a lost rotor from the residual between commanded and measured rotor speed. Then give up yaw control: the vehicle spins about its yaw axis and controls its reduced attitude.
Implement: a mixer mode plus an attitude-loop mode.
Test: SITL fault injection (racing-drone#16 adds fault injection).
Strata: S0 – · S1 – · S2 ● · S3 ○ · S4 ○ · S5 –

B. Rate loop

B1. Cascaded PID rate loop — reflex#20.

Design:
- D on measurement, with a D-term low-pass filter.
- Integrator anti-windup driven by the mixer’s saturation report (A1).
- Feed-forward from the setpoint derivative.
- “I-term relax” during fast stick or setpoint moves.
Implement: control/pid.hpp (exists) and control/rate_controller.hpp.
Test:
- Bandwidth and overshoot specifications per stratum parameter set.
- No integrator windup through a 2 s saturation.
Strata: ● in every stratum.

B2. Gyro filtering — reflex#21.

Design:
- Low-pass filter plus static and dynamic notch filters.
- RPM-tracked notches at the rotor harmonics.
- The filter-delay budget is part of the phase margin.
Implement: dsp/.
Test:
- Group delay at the crossover frequency below a specified limit.
- Attenuation at the rotor frequency.
- SITL with injected vibration.
Strata: S0 ○ · S1 ● · S2 ○ · S3 ● · S4 ● · S5 ●

B3. INDI (incremental nonlinear dynamic inversion) as an alternative rate loop.

Design:
- Angular-acceleration estimate from the filtered gyro.
- Control effectiveness from the mixer matrix.
- Rotor-speed feedback.
- Robust to model error because it acts incrementally.
Implement: control/indi.hpp.
Test:
- Disturbance rejection against the PID baseline in SITL.
- Sensitivity to errors in the control-effectiveness estimate.
Strata: S0 – · S1 ○ · S2 ○ · S3 ● · S4 ● · S5 ●

B4. Gain scheduling.

Design:
- Throttle-PID attenuation (TPA) at high collective thrust.
- At speed, schedule on dynamic pressure $q$ , with the estimate of $q$ coming from cortex.
- Gains are fixed per schedule point: no reconfiguration at runtime.
Implement: a schedule table in the rate controller.
Test:
- Stability margins at every schedule point.
- Bumpless transfer between points.
Strata: S0 – · S1 – · S2 ○ · S3 ● · S4 ● · S5 ●

C. Attitude loop

C1. Quaternion attitude control on $SO(3)$ — reflex#20.

Design:
- No Euler angles anywhere in the control path.
- Error from $q_{err} = q^{-1} \otimes q_{sp}$ , taking the shortest-path sign.
Implement: control/attitude_controller.hpp.
Test:
- Large-angle steps of 180° in roll and pitch.
- Hold at 85° body tilt (the body z axis 85° from world vertical), as an envelope stress case. On an uncanted airframe body tilt and rotor-thrust tilt coincide; on a canted S5 airframe the test applies to the body attitude envelope, set per vehicle under H3.
- Continuity of the error through the antipodal quaternion.
Strata: S0 ○ · S1 ● · S2 ○ · S3 ● · S4 ● · S5 ●

C2. Tilt-prioritized (reduced-attitude) control.

Design:
- Split the error into tilt and yaw.
- Track the tilt first, because it carries the thrust vector.
- Yaw has lower bandwidth and yields under saturation, consistent with the mixer priority (A1).
Implement: the attitude controller.
Test:
- Combined tilt and yaw steps.
- The tilt error converges regardless of the yaw error.
Strata: S0 ○ · S1 ● · S2 ○ · S3 ● · S4 ● · S5 ●

D. Translational control and the planner interface

D1. CTBR command interface.

Design:
- Collective thrust plus body rates, with optional torque feed-forward.
- The command carries a timestamp, a validity window, and staleness handling.
Implement: control/ message types, mirrored in racing-drone’s msg/.
Test:
- Staleness: a command older than the window is replaced by the safe fallback.
- The latency is measured end to end.
Strata: S0 ○ · S1 ○ · S2 ○ · S3 ● · S4 ● · S5 ●

D2. Velocity and position loops — reflex#20.

Design:
- Acceleration command → thrust vector → attitude setpoint.
- Feed-forward from differential flatness.
- Drag compensation.
- Thrust-vector saturation that keeps the vertical component first.
Implement: control/position_controller.hpp.
Test:
- Racing-drone gate P1: ≥ 99 % of 500 seeds at ≤ 5 m/s over two laps.
- Tracking RMS per stratum.
Strata: S0 ● · S1 ● · S2 ● · S3 ● · S4 ○ · S5 ○

D3. MPC (MPCC) as the planner.

Design: RTI-SQP, with the model in E as the prediction model.
Implement: reflex roadmap Phase 3.
Test: racing-drone gates P2–P3.
Strata: S0 – · S1 – · S2 – · S3 ● · S4 ○ · S5 –

E. Aerodynamic model terms

These are needed in QuadrotorModel for SITL fidelity, and in the controller for compensation.

E1. Rotor drag (blade flapping and H-force).

Design: a force proportional to the in-plane velocity, scaled by thrust: $-k_d\, T\, v_{\perp}$ .
Test: compare the predicted acceleration against flight logs, using MonoRace’s identified terms [5] as a reference.
Strata: S0 – · S1 ○ · S2 ○ · S3 ● · S4 ● · S5 ●

E2. Thrust lapse $C_T(J)$ .

Design: a thrust coefficient that depends on the axial advance ratio, from propeller data or a blade-element model.
Test: thrust-stand data or a wind-tunnel table.
Strata: S0 – · S1 – · S2 ○ · S3 ● · S4 ● · S5 ●

E3. Body lift and moments.

Design: lift, drag, and pitching moment as functions of the angle of attack, from AirShaper-style CFD or a wind tunnel.
Test: CFD tables.
Strata: S0 – · S1 – · S2 – · S3 ○ · S4 ● · S5 ●

E4. Ground and ceiling effect.

Design: thrust gain as a function of height over rotor radius.
Test: a hover-height sweep.
Strata: S0 ● · S1 ● · S2 ○ · S3 ○ · S4 – · S5 –

E5. Vortex ring state in fast descent.

Design: detect it, and constrain the vertical descent rate.
Test: an SITL descent-rate sweep.
Strata: S0 ○ · S1 ○ · S2 ● · S3 ○ · S4 – · S5 –

E6. Compressibility at the prop tips.

Design: correct $C_T$ and $C_Q$ for the helical tip Mach number.
Test: propeller data.
Strata: S0 – · S1 – · S2 – · S3 – · S4 ○ · S5 ●

F. Interfaces to estimation

F1. IMU saturation and model fallback.

Design:
- Flag accelerometer and gyro saturation.
- When saturated, the controller uses the model-predicted acceleration, as MonoRace does [5].
- Shared with cortex.
Test: SITL with sensor clipping injected.
Strata: S0 – · S1 ○ · S2 – · S3 ● · S4 ● · S5 ●

F2. Latency compensation.

Design:
- Forward-propagate the estimate to the actuation time with IMU propagation (reflex#13 roadmap).
- The budget is stated per stratum.
Test: delay injection in SITL. The tracking error must stay within specification at 1.5× the nominal delay.
Strata: S0 ○ · S1 ○ · S2 ○ · S3 ● · S4 ● · S5 ●

F3. Vibration aliasing.

Design:
- Choose the IMU sample rate and anti-alias filtering against the rotor frequencies.
- Notch frequencies follow the RPM (B2).
Test: spectral tests on logs.
Strata: S0 ○ · S1 ● · S2 ○ · S3 ● · S4 ● · S5 ●

G. Power, thermal, and structure

G1. Battery and current limits.

Design:
- Make the thrust envelope a function of state of charge and current limit.
- The mixer’s maximum thrust comes from the power stage, not a constant.
Test: an SITL full-pack run with sag.
Strata: S0 ● · S1 ● · S2 ○ · S3 ● · S4 ● · S5 ●

G2. Thermal derating.

Design: derate the ESC and motor current limits with temperature, and model the temperature itself.
Test: thermal-model SITL, then a bench test.
Strata: S0 – · S1 – · S2 – · S3 ○ · S4 ● · S5 ●

G3. Aeroelastic and structural modes.

Design: frame resonances and flutter speed; the structure must keep a margin above the control bandwidth.
Test: modal survey and notch placement.
Strata: S0 – · S1 ○ · S2 – · S3 ○ · S4 ● · S5 ●

H. Safety and lifecycle

H1. Arming, disarming, and idle.

Design:
- The airmode thrust shift (A1) is active only when armed.
- Motor idle is set on the ground.
- Disarm on landing detection.
Implement: the FC state machine.
Test:
- State-machine model checking.
- No motor output while disarmed.
Strata: ● in every stratum.

H2. Failsafes.

Design:
- Link loss → hover, land, or kill, depending on the stratum.
- Stale-command handling (D1).
- Estimator-health gating with NIS (reflex#13).
Test: fault-injection scenarios.
Strata: ● in every stratum.

H3. Envelope protection.

Design: limits on tilt, body rate, and speed that are configurable per stratum, plus a geofence.
Test: SITL scenarios commanded beyond the envelope.
Strata: S0 ● · S1 ● · S2 ● · S3 ○ · S4 ● · S5 ●

Test strategy across strata

Strata make a natural test matrix. The proposal:

One vehicle parameter set per stratum, as racing-drone vehicles/*.yaml files with reflex test fixtures that mirror them.
- S3 is calibrated from the MonoRace and RP-1 numbers.
- The other strata start from the reference values in the summary table and are labelled as estimates until measured.
Component tests run across the sweep. The mixer, rate, and attitude tests are parameterized over S0–S5, in the same way that Pid is parameterized over scalar types.
- Specifications are written in normalized form: bandwidth as a fraction of the motor bandwidth $1/\tau$ , overshoot in percent, settling time in multiples of $\tau$ .
- One test can then judge every stratum.
SITL scenarios per stratum.
- Hover and position hold for all strata.
- Gate courses for S3.
- Straight-line speed runs with $C_T(J)$ enabled for S4 and S5, once E2 lands.
- Each scenario reports a scorecard: tracking RMS, percentage of time saturated, peak tilt, energy used.
Model validity gates. A SITL result is reported for a stratum only when the model includes every term that stratum marks ● in section E. Otherwise the scorecard is flagged model-limited.
Hardware validation.
- Bench: thrust stand, then HIL on the flight controller.
- Flight: S3 is the program’s flight target (RP-1). The other strata are validated in simulation unless a vehicle becomes available.

Recommended order of work

A1 mixer, then B1/C1/D2 cascaded PID, then B2 filters. This is the racing-drone Phase 1 path to gate P1, which flies S3 hardware slowly.
A2 and A3 (thrust linearization and motor lag), D1 (the CTBR interface), and F1/F2. These are needed before S3 flies fast.
E1 rotor drag and E2 thrust lapse in the model, then the S3 SITL gates at speed.
B3 INDI and B4 gain scheduling as the high-speed rate loop.
E3, E6, G2, G3, only if S4 or S5 become program targets.

References

Bitcraze, Crazyflie 2.1 datasheet — https://www.bitcraze.io/documentation/hardware/crazyflie_2_1/crazyflie_2_1-datasheet.pdf
Bitcraze, Crazyflie 2.1 Brushless — https://bitcraze.io/products/crazyflie-2-1-brushless
GetFPV, BetaFPV Meteor65 1S Brushless Whoop — https://www.getfpv.com/betafpv-meteor65-1s-brushless-whoop-quadcopter-22000kv.html
DJI, Mini 4 Pro specifications — https://www.dji.com/mini-4-pro/specs
TU Delft MAVLab, MonoRace (arXiv 2601.15222): 966 g, 5.1” props, τ = 25 ms, 28.23 m/s, ≈ 7 g, 16 g IMU limit — https://arxiv.org/abs/2601.15222
Drone Champions League, The drones (850–1000 g, ≥ 360 mm diagonal, > 160 km/h) — https://dcl.aero/explore/drones/
UAVModel, Best FPV drone motors — thrust testing; MAD BSC 2207.5 listing (1.85 kg) — https://blog.uavmodel.com/6111-2/
DroneDJ, Red Bull drone chases F1 (350 km/h, < 1 kg, 42 000 rpm, ≈ 3 min) — https://dronedj.com/2024/02/29/red-bull-drone-f1-fastest/; Newsshooter — https://www.newsshooter.com/2024/02/27/worlds-fastest-camera-drone-300-km-h-in-just-4-seconds-with-a-top-speed-of-over-350-km-h/
LIGA.net, Peregreen 3 reaches 585 km/h (2.77 kg, 16.2 kW, water cooling) — https://tech.liga.net/en/technology/news/drone-printed-on-a-3d-printer-reached-speed-of-585-kmh---video
New Atlas, Peregreen V4 reclaims speed record (T-Motor 3120 900 KV, 6” props) — https://newatlas.com/drones/luke-mike-bell-peregreen-v4-guinness-speed/
Inspenet, Peregreen 2 Guinness record (Velox 3115 900 KV, 12S, 120 A ESCs) — https://inspenet.com/en/news/quadcopter-record-guinness-speed-recorder/; Guinness World Records, 480.23 km/h — https://www.guinnessworldrecords.com/news/2024/5/father-and-son-duo-create-worlds-fastest-drone-that-reaches-500-km-h
Tom’s Hardware, Peregreen V4 at 657.59 km/h — https://www.tomshardware.com/3d-printing/fully-3d-printed-drone-capable-of-flying-at-408-mph-is-the-fastest-in-the-world-father-son-duo-reclaim-guinness-world-record-title-with-diy-quadcopter
Data-Driven System Identification of Quadrotors Subject to Motor Delays (Crazyflie τ ≈ 0.072 s; earlier estimate 0.065 s) — https://arxiv.org/abs/2404.07837
Bitcraze, PWM to thrust and Keeping thrust consistent as the battery drains — https://www.bitcraze.io/documentation/repository/crazyflie-firmware/master/functional-areas/pwm-to-thrust/
Abu Dhabi Autonomous Racing League (A2RL drones > 140 km/h) — https://en.wikipedia.org/wiki/Abu_Dhabi_Autonomous_Racing_League
Notebookcheck, DIY engineers 3D-print world’s fastest drone (Peregreen 3: 16.2 kW peak draw, 2.77 kg, extreme-pitch APC props against supersonic tips; Peregreen 2: 8.9 kW, 1.85 kg) — https://www.notebookcheck.net/DIY-engineers-3D-print-world-s-fastest-drone-with-363-5-mph-top-speed.1146428.0.html