Hardware combinations, ranked by evidence
Three research agents went looking for the cheapest hardware that genuinely balances a cube. The interesting results were the negative ones: the most-recommended motor in this hobby fails half the spec, the most-recommended cheap driver cannot do torque control at all, and the cleverest salvage idea is 10 to 100 times too weak.
The constraint everyone states as one number [ it is two ]
A reaction wheel for a 100–120 mm cube needs roughly 0.08–0.13 N·m of torque. That number gets quoted constantly. What gets dropped is the other half: the wheel also has to reach 1500–2000 rpm, because stored angular momentum is what actually arrests a fall, and momentum is inertia times speed.
Those two requirements pull in opposite directions in a motor winding, and that is where the popular advice quietly breaks:
The iPower GBM4108H-120T is the motor this hobby recommends most, and it fails the speed half. It makes the torque comfortably — roughly 0.12–0.18 N·m — but its own datasheet quotes 513–567 rpm at 20 V, implying about 26–28 Kv. It physically cannot reach 1500 rpm at any hobby bus voltage. Great torque, wrong tool.
Of everything checked, exactly one part verifiably clears both halves under $70: the CubeMars GL35 KV100 — 0.15 N·m continuous, 0.46 N·m peak, 90 g, $54.99 on the vendor's own page. Even that is borderline on speed at 3–4S (~1200–1650 rpm no-load), so bench-verify loaded speed before committing to a flywheel inertia.
Four costed stacks [ vendor-checked where marked ]
A — Rock bottom, one axis, ~$58
GBM2804H-100T or a generic 2204/2208 gimbal motor ($15–20) · SimpleFOCMini ($6.59) · AS5600 ($4) · MPU6050 ($2) · STM32F411 blackpill ($6) · 3S 1300 mAh ($15) · bearing, standoffs, printed parts (~$9).
Can: prove the concept on one edge with a cascaded angle/velocity loop — architecturally this is SimpleFOC's own reference pendulum. Cannot: true torque control (the Mini is voltage-mode), three axes, or firm disturbance rejection. The 2804's ~0.05–0.098 N·m sits below the target band, so expect it to be tippy under a real push.
B — Best value, three axes, ~$419
3× CubeMars GL35 KV100 ($165) · 3× B‑G431B‑ESC1 (~$88–117) · 3× AS5047P (~$51) · ICM‑42688‑P (~$15) · ESP32‑S3 DevKitC ($15) · 4S 1500 mAh ($22) · buck regulator ($3) · mechanical (~$60).
Can: genuine corner balancing with real per-axis torque control — each G431 closes its own current loop while the S3 runs fusion and the balance law. Cannot: be treated as a kit. No single library glues three independent ESC boards to one coordinator; that firmware is yours to write.
C — Premium, ~$505
Same architecture, upgraded sensing: MT6835 21-bit encoders and a Teensy 4.1 coordinator with room for logging and telemetry, plus bus-voltage headroom for speed margin. Buys you diagnosability, not capability.
D — Wildcard: gut a broken camera gimbal, ~$55–100
A "for parts" 3-axis brushless gimbal off eBay (from about $26) hands you three mechanically matched, purpose-built gimbal motors, sometimes with encoders already fitted. The catch: its own controller (SimpleBGC/AlexMos/Storm32) is built for stabilisation holding, not for accepting an external torque command, so you rebuild the drive electronics regardless — and you will not know the motors' Kv until they are in your hands, with a real risk of landing exactly the GBM4108 problem above.
Clever ideas, honestly assessed [ mostly negative results ]
| Idea | Verdict | The numbers |
|---|---|---|
| Hard-drive spindle motor as the reaction wheel | Won't work | Kt ~4–6 mN·m/A, and the windings are optimised for low current at high speed, not stall torque. Estimated peak ~10 mN·m against a 50–130 mN·m requirement — short by 5 to 100×. The CubeSat precedent that makes it look plausible was built for disturbance torques around 0.0006 mN·m. |
| HDD platters as the flywheel | Viable — the good half of that instinct | A 3–4 platter stack computes to roughly 0.8–1.1×10⁻⁴ kg·m², right alongside a purpose-built hobby flywheel measured at 1.36×10⁻⁴. Free, precision-balanced, already round. Bin the spindle motor, keep the discs, print a hub. |
| Racing FPV motors (2205/2207) | Won't work | Kt ≈ 9.55/Kv ≈ 0.004 N·m/A on a 2300Kv RS2205, so 0.1 N·m needs ~25 A continuous through windings cooled by prop wash that isn't there. The SimpleFOC community tried it: "excessive speed with poor fine control." |
| Cheap 4-in-1 ESC + AM32/Bluejay firmware | Won't work as a torque source | These firmwares are throttle/speed oriented; DShot commands something closer to voltage than to current. The one documented reaction-wheel attempt used ESC current draw as a torque proxy and the builder called it "not exactly correct." Sensorless BEMF also goes blind at the zero crossings a balancing wheel lives on. |
| Hoverboard motor + hoverboard-firmware-hack-FOC | Viable — at the wrong scale | ~10–15 N·m for a $50–100 donor, i.e. ~100× more than a desk cube needs, at several kg per motor. Genuinely excellent mature open-source FOC if you are building something human-sized. |
| Skip the encoder (sensorless FOC) | Worst possible fit | Back-EMF amplitude scales with speed, so position estimation collapses near zero — and a balancing wheel deliberately reverses through zero constantly. You would be saving a $2–5 part to break the one thing you need. |
| Open-loop stepper reaction wheel | Works, with a real ceiling | ~$7 for a NEMA17 plus DRV8825, no encoder at all, and it is a documented working attitude-control build. Position is known without feedback as long as you never lose steps. Costs you top speed (so, stored momentum), adds resonance, and worsens torque-to-mass. |
| CyberGear / Steadywin quadruped actuators | Wrong regime | Remarkable value — 4 N·m continuous, integrated FOC and encoder, ~$70 — but the 7.75:1 gearbox caps output at 296 rpm. Torque you don't need, speed you can't do without. Better suited to a CMG gimbal axis. |
| Control moment gyroscope instead of a reaction wheel | Real, but not cheaper here | CMGs win enormously on torque per watt at spacecraft scale. At desk scale the extra gimbal motor, extra encoder and singularity-avoidance steering law cost more in money and time than the reaction wheel's simplicity saves. |
| Ducted fans / cold gas | Won't work | A reaction wheel draws power only while changing momentum; a fan burns power just to hold a torque, and adds spin-up lag. Telling detail: even the Cubli's jump mechanism is a servo-actuated mechanical brake, not a gas jet. |
Silicon: three traps [ all from vendor docs ]
- ESP32-C3 has no hardware FPU — it does software floating point, and it lacks the MCPWM peripheral entirely (only LEDC), which cannot sync with ADC sampling for low-side current sensing. Use a plain ESP32 or an S3.
- RP2040 has no FPU either, and SimpleFOC's own docs state its ADC "features are not sufficient for low-side current sensing," with inline support "recently developed, and not well tested." RP2350 fixes the FPU; the motor-control timer hardware is still not G431-class.
- BNO055 caps fusion output at 100 Hz — below the floor for this problem. If you use a fusion IMU, read its raw gyro and accelerometer, not its fused output. The best noise-and-speed part found was the ICM‑42688‑P at 2.8 mdps/√Hz and 32 kHz ODR.
The ranking [ if you only read one thing ]
- Gimbal BLDC + a real current-sensing FOC board + magnetic encoder + decent IMU. Boring, ~$45–55 per axis, and it is what essentially every documented working build actually used. Everything else on this page is a variation on it or a cautionary tale.
- The same, but with salvaged HDD platters as the flywheel mass. The one genuinely clever cost cut that survives contact with the numbers.
- B‑G431B‑ESC1 as an all-in-one node. Costs more than a Mini plus a blue pill, removes the entire current-sensing class of problem, and scales to three axes by buying three.
- Scavenged gimbal-class motor — cheap, provided you resist re-flashing the ESC and drive it properly instead.
- Hoverboard stack, only if the thing you are balancing is human-sized.
Confidence, stated plainly. Prices and specs here were pulled from vendor and datasheet pages on 17 September 2026, and are presented as verified only where a page was read directly. Two independent checks returned different DigiKey prices for the B‑G431B‑ESC1 ($29.25 and $38.75), so treat that line as $29–54 until you see a cart total. Every "mechanical" line — frames, flywheels, bearings — is an estimate with no vendor quote behind it, and is the least reliable number on this page. Torque figures for generic 2204/2208 motors are extrapolations, not datasheet values.