Out in the yard, with the aircraft just sitting there, everything looks perfect. The motors arm, GPS has a fix, the compass shows no deviation. Then, halfway through a full-throttle climb, the telemetry link drops for a second, an "EKF variance" warning appears in the log, and the companion computer reboots. You go straight to the software; nothing in the code changed.

Most failures of this kind trace back to power distribution and cable routing. On a multirotor, tens of amps are being switched between the battery and the motors, and that current creates both a voltage drop along the wires and a changing magnetic field around them. If the compass sits 10 cm away, it reads that magnetic field; if the 5 V rail sags by 0.3 V, the Jetson shuts down.

We build the power architecture around four questions: which load gets fed from which rail, what gauge the wire has to be, where the noise gets in, and how much headroom to leave at peak current. On our SUAS 2026 aircraft (Pixhawk-class FC running ArduPilot, Jetson Orin NX, SIYI A8 mini gimbal, servo-based airdrop) we rewrote those answers several times.

Power architecture: separate rails, not one bus

The main chain is simple: battery → PDB (power distribution board) → ESCs. The PDB's only job is to split the high current coming off the battery among the motor drivers. The hard part is how you branch the "clean" rails off that main bus with voltage regulators.

ESCs switch current at high frequency and leave ripple on the main bus. If you feed both a servo and the FC from the same regulator, the current spike that occurs the moment the servo moves shows up on the FC's 5 V rail. That is why we split loads by their character.

ArduPilot's Pixhawk power documentation sets a hard limit: in normal operation the FC power input and the servo rail accept 4.1–5.7 V. Digital servos can push the rail above 5.7 V, which makes the power management circuit shut the board down and restart it. If you plan to use the servo rail as a backup supply, clamp it with the zener diode the documentation recommends (1N5339).

Choosing a BEC and sizing current headroom

Two things matter when you pick a BEC (battery eliminator circuit — the step-down converter that provides a regulated output): continuous current capability, and how much of the loss ends up as heat.

Switching BEC:

P_out = V_out × I_out

P_in = P_out / η (η: efficiency, typically 80-90% — check the manufacturer's data)

P_loss = P_in - P_out (the part that turns into heat inside the BEC)

Linear regulator:

P_loss = (V_in - V_out) × I_out

An example: a load drawing 5 V / 4 A. With a switching BEC, P_out = 20 W, and at η = 0.85 that gives P_in = 23.5 W with 3.5 W of loss. Feed the same load from a 6S battery (22.2 V nominal) through a linear regulator and the loss becomes (22.2 − 5) × 4 = 68.8 W. That is why linear regulators are only used for loads of a few hundred milliamps.

Our headroom rule: I_BEC ≥ 1.5 × Σ(peak currents). By peak we mean the stall (locked-rotor) current from the datasheet, not the servo's idle draw. For the BEC feeding the servo rail, the ArduPilot documentation recommends leaving 2-3 A extra for the FC, the receiver, and brownout margin.

Wire gauge: running the AWG numbers

You do not guess at wire size, you calculate it in two steps: current capacity first, then voltage drop.

R_line = 2 × L × r_awg (total out-and-back resistance, Ω)

V_drop = I × R_line (V)

P_loss = I² × R_line (W)

L: one-way cable length (m)

r_awg: resistance per unit length (Ω/m), from the table

I: current through the run (A)

The "chassis wiring" column applies to short runs in open air that are not inside a loom; if you bury the wire in the middle of a bundle, derate below those figures.

Example 1 — the main bus. 6S pack, 40 A total in hover, 15 cm between battery and PDB, 12 AWG:
R_line = 2 × 0.15 × 0.00521 = 0.00156 Ω → V_drop = 40 × 0.00156 = 0.063 V (0.3% of 22.2 V), P_loss = 40² × 0.00156 = 2.5 W. Pull 120 A at full throttle and the loss in that same wire climbs to 22.5 W; it is short-lived, but you can feel the wire warm up.

Example 2 — a regulated rail, the real trap. Say we draw 4 A from a 5 V BEC through 1 meter of 20 AWG:
R_line = 2 × 1.0 × 0.03329 = 0.0666 Ω → V_drop = 4 × 0.0666 = 0.27 V. Only 4.73 V reaches the Jetson's input. NVIDIA specifies a VDD_IN range of 5–20 V for the Orin NX module, so 4.73 V is already below the limit. Move that same run to 16 AWG and the drop falls to 0.105 V.

The better fix: carry the same power at a higher voltage. If your carrier board accepts 12 V, 25 W is about 2.1 A, the drop over 1 meter of 20 AWG is 0.14 V, and against 12 V that is negligible. Always confirm your own carrier board's input range against the manufacturer's documentation; the module's range and the board's range are not the same thing.

Connectors and soldering

The number in a connector's name is not its continuous rating. Here are the original AMASS figures as published by Holybro:

If your hover current is above 30 A, you are sitting right at the XT60's continuous limit — move to XT90. On 6S and larger packs, use the anti-spark version (XT90-S) to handle the arc you get on plug-in.

Three rules for soldering: use an iron with enough power for heavy wire (a small iron melts the solder but never heats the copper, and the result is a cold joint), wick the solder into the strands, and insulate with heat-shrink. A cold joint is a high-resistance joint; it heats up in flight, its resistance rises, and then it fails.

Safety. A LiPo pack can deliver hundreds of amps into a short — the wire melts within seconds and starts a fire. The battery must never be connected while you are soldering. Insulate bare XT60 terminals individually, and never grab both at once with pliers. Do not fit props while the motors are wired up; keep the props off during testing.

Capacitors: soaking up voltage spikes

ESCs chop the current on and off quickly. The inductance of the wire running back to the battery turns each of those interruptions into a voltage spike, and the longer the wire, the bigger the spike. Sources that have actually measured it report peaks at the battery terminals of a 6S system reaching nearly twice the average voltage under active braking.

The fix is to solder a low-ESR (equivalent series resistance) electrolytic capacitor across the ESC's battery input. Practical rules:

• If the battery leads are longer than 15 cm, increase the capacitance (1000 µF is a common starting point on 5-inch-class 6S builds).

• The voltage rating has to sit above the pack's fully charged voltage: 6S means 25.2 V when full, so 35 V is marginal and 50 V leaves comfortable margin.

• Keep the capacitor's leads as short as you can; long leads add inductance of their own and blunt the filtering.

• Do not reverse the polarity. An electrolytic capacitor connected backwards will explode.

EMI: where the noise gets in

Magnetic interference is the compass mistaking motor current for magnetic north. The ArduPilot documentation gives clear prescriptions for it:

• Mount the GPS/compass module at least 10 cm away from DC power cables and the battery, on a mast, with a clear view of the sky.

• Keep the runs between battery, PDB and ESCs as short as possible and twist them together (twisted pair). Twisting the outbound and return conductors cancels most of the field against itself.

• Use grounded shielding where you can; wrapping the ESC-to-motor leads in aluminum shield can reduce AC interference.

• 4-in-1 ESCs generally produce less interference, because their wiring is short and kept together.

• If a CompassMot test shows interference above 60%, change the layout — do not try to calibrate your way out of it.

The same logic applies to the Ethernet/USB cables carrying telemetry and RTSP: do not run them parallel to power cables over long distances, and where they have to cross, cross at right angles. A ferrite bead cuts high-frequency common-mode noise, but it will not fix a layout mistake; get the spacing right first and add the ferrite as a finishing touch.

Companion computer and brownout risk

A companion computer like the Jetson is not a constant load. NVIDIA defines 10 W / 15 W / 25 W power modes for the Orin NX 16GB, and JetPack 6.2 added a 40 W reference mode plus an uncapped MAXN SUPER mode. The moment model inference starts, GPU draw jumps to several times its idle figure.

So build the power budget around peak, not average. In 25 W mode the 5 V rail sees roughly 5 A; a BEC labeled "5 V 5 A" runs right at its limit and shuts down once it gets hot. Decide which power mode you are going to run (that choice also drives FPS — our measurements are in Tiling or speed) and add 50% on top of that mode's peak.

Here is our budget: the SIYI A8 mini gimbal camera runs on 11–25.2 V, drawing 5 W on average and 12 W at peak. The Jetson has its own regulator; the FC and GPS are fed through the power module. The airdrop servo sits on a separate BEC — on missions where we release from about 20 m, it mattered that servo motion never touched the FC rail. Details are in the Jetson Orin NX setup guide and the companion computer MAVLink guide.

Catch brownouts on the bench: run the aircraft with the props off, start a heavy inference loop on the Jetson, drive the servos at the same time, and watch the POWR (power status) records in the FC log. The moment the rail collapses shows up here — not in the air.

Common traps

• Hanging everything off one BEC. One device's peak current drags down the rail for all the others.

• Thin wire. A long run of thin wire on a 5 V rail is the most common cause of brownouts. The 0.27 V drop in Example 2 is invisible, and lethal.

• Putting the GPS next to a power run. It looks fine on the ground; once the motors load up, the compass drifts and the aircraft starts flying circles.

• Reading the connector label as a continuous rating. An XT60 is good for 30 A continuous; 60 A is the burst figure.

• Skipping the capacitor, or soldering it in with long leads. Long leads stop the capacitor from doing its job.

• Budgeting against average draw. Peak current is brief, but it is long enough to pull the rail down.

Practical summary

1. Separate the rails: FC/GPS from the power module, the servo from its own BEC, the companion from another BEC, the camera from its own feed. Keep the FC power input inside 4.1–5.7 V.

2. Do the wire math: V_drop = I × 2 × L × r_awg. On regulated rails keep the drop below 2% of nominal voltage; if you cannot, carry the power at a higher voltage and regulate it at the load.

3. Pick connectors by continuous current: above 30 A in hover move to XT90, above 45 A to the XT120 class. Solder with a heavy iron, and never solder with the battery connected.

4. Put a low-ESR capacitor on the ESC input, leads short, voltage rating above the fully charged pack voltage.

5. Solve EMI with layout: GPS/compass at least 10 cm away from power cables, power runs short and twisted, CompassMot below 60%.

6. Budget for peak current and load-test on the bench; do not discover a brownout in the air.