Bill of materials

Prices are estimates. EUR, ex-shipping, mid-2026 European availability. They're good enough to budget from, not to reconcile accounts against. Lines marked ❌ are the only things not yet on hand.

These are the 19 kg·cm / 7.4 V servos, not the 30 kg·cm / 12 V ones. Torque is roughly a third lower than the 12 V variant most SO-101 build logs assume. Gravity sag is more visible and payload is lower. Keep anything bolted to the wrist under ~40 g, and expect shoulder_lift to work hard at full extension.

Crowbar the 7.4 V rail. The gap between safe (7.4 V) and destructive (12 V) is 4.6 V. A knocked trimpot costs you twelve servos. Four mitigations, all of them:

1. Set the output and verify with a multimeter into a dummy load before any servo is connected.
2. Lock the trimpot — nail varnish or thread-lock over the screw.
3. Fit the 8.0 V TVS across the rail, downstream of the 10 A fuse. If the output rises, the TVS conducts hard and blows the fuse before the servos see it.
4. Set the converter's CC limit to ~10 A so a bus fault current-limits instead of running away.

The XY6020L retires 1, 2 and 4 as manual steps: the voltage and the current limit are both keyed in and stored, and its display makes verification continuous rather than a one-off multimeter check. There is no pot to varnish because there is no pot.

3 still applies, unchanged. A digital setpoint is not a fault mode — a shorted high-side FET passes raw pack voltage to the servos regardless of what the display says. The TVS and the fast-blow fuse are the only mitigations that don't depend on the converter working, which is exactly why they're the ones you can't skip.

Two things to confirm on arrival: that the module's output-on-at-power-up setting is enabled, so the arms come back without a button press after an E-STOP cycle, and that it has real airflow — 24 V → 7.4 V at 10 A is 74 W out.

Correction: P6KE8.2A is the wrong TVS and earlier revisions of this page listed it. It would conduct on a healthy rail.

The two series number their parts differently. SMBJ8.0A's number is the stand-off voltage (8.0 V, breakdown 8.89–9.83 V). P6KE8.2A's number is the breakdown voltage — its stand-off is only ~7.02 V, below the 7.4 V rail, so it sits in leakage the moment the rail comes up, and its 7.79 V minimum breakdown is close enough to nominal to nuisance-clear the 10 A fuse on ripple.

The through-hole part matching SMBJ8.0A's intent is P6KE9.1A (stand-off ≈7.78 V, breakdown 8.65–9.55 V). Confirm V_RWM against your specific manufacturer's datasheet before fitting.

6 A blade fuses don't exist. The standard ATO/ATC series is 1, 2, 3, 4, 5, 7.5, 10, 15, 20, 25, 30, 35, 40 A. The per-arm branches specified as 6 A should be 7.5 A — a 5 A blade sits exactly at each arm's ~5 A peak and will nuisance-blow mid-episode.

The tradeoff to make deliberately: 7.5 A is marginally above the 18 AWG pigtails' stated 7 A rating, so the fuse no longer strictly protects the wire. 18 AWG silicone in free air carries well beyond 7 A and that figure is a connector rating rather than a conductor limit — but make the call knowingly.

Voltage drop matters far more at 7.4 V than at 12 V. The servo floor is 6.0 V, so you have 1.4 V of headroom for the whole chain. Use 14 AWG for the 7.4 V run, keep it under 0.5 m, and mount the converter near the arms.

DC5521 barrel jacks are rated ~5 A — exactly where each arm sits at peak. They warm up. For a permanent build, solder power directly to the Waveshare board's DC input pads rather than relying on the barrel.

Don't buy the 5 A version of the 12 V converter. The rail carries the Jetson at 3.75 A plus the powered hub at ~1 A plus the E-stop relay coil — about 4.8 A. A 5 A part would sit at ~96% of rating continuously, which is a thermal-shutdown design rather than a margin. The 10 A class costs a few euro more. It doesn't enlarge the fault energy either: the 5 A fuse on the converter's input still bounds that branch.

The bandwidth trap. Two 640×480@30 cameras in raw YUYV need ~147 Mbps each; USB 2.0 delivers ~280 Mbps practical. Put both on one root and you get VIDIOC_STREAMON: No space left on device. Buy cameras with MJPEG and force MJPEG in the capture config — roughly 10× less bandwidth.

wrist_roll is calibrated as continuous. A USB cable routed through it will wind up and snap. Clamp the range in software or leave a generous service loop with strain relief at the wrist.

Don't run USB alongside servo power for long parallel stretches. The bus is a half-duplex 1 Mbaud line sitting next to a switching converter.

You can start recording before the cameras arrive. The D455 alone, rigidly mounted on the tower, is a complete single-camera observation. That path needs none of the €118 below the blocking line — see dataset collection.