All work

Project Ferros · In development

Tools that build tools.

— the bench should hold the board, not you

Ferros watches where your soldering iron lands, waits for the joint to actually reach temperature, and feeds solder to that exact point. No foot pedal. No third hand you don't have.

180° ARCH TRACK
01

Parked on the rail

The arm has no base on the bench. It hangs from a carriage that rides the arch itself, and at rest it parks near one foot — folded up, clear of the board, out of your way.

02

180° of travel

A pinion on the carriage drives a rack cut into the arch, carrying the arm from one foot up over the apex to the other. The depth camera and thermal array sit on a mast above the apex, clear of the carriage's path; an RGB camera on a standoff beside each foot covers the work from its own angle, so a hand can't hide the joint from every view at once.

03

Double-tap to designate

Tap the joint twice with the iron tip. A piezo disc under the fixture catches both impulses, the camera confirms the tip was on the board, and that pad becomes the target.

04

Solder, on the joint

The thermal array watches the pad. The moment its trace shows the melt signature, the feeder advances a metered length of wire — then retracts 2–3 mm to break the bridge.

The problem

Soldering needs three hands. You have two.

One hand holds the iron. One holds the board. The solder gets whatever's left.

The technique every tutorial teaches — iron heats the joint from one side, solder feeds into the joint from the opposite side — is exactly the part that gets sacrificed. Solder fed onto the tip instead of the joint loses its flux before it reaches the work, and you get the cold, grainy joint.

Helping-hands clamps and manual solder feeders exist, but they're static and human-triggered: you position them once and work around them, then press a pedal. For someone with the use of one arm, those workarounds collapse entirely. Ferros moves to you instead, and decides when to feed on its own.

The hard part

Molten solder lies to a thermal camera.

Infrared thermometry measures radiance and infers temperature from an assumed emissivity. Board material sits near 0.9 — easy. But the instant solder melts it becomes a shiny specular metal, and its emissivity collapses toward 0.05–0.15. The apparent temperature dips at the exact moment the real temperature is rising.

So a naive threshold — "feed at 190 °C" — mistimes every joint. The fix is to treat that artifact as the signal: the detector looks for the latent-heat plateau, the emissivity dip, and the adjacent board pixels rising smoothly past a calibrated band. A taped-on K-type thermocouple provides ground truth to calibrate it against.

Hardware

What's inside.

Depth
Luxonis OAK-D PoE

Stereo depth with on-device inference, looking down from the arch apex. Board-frame coordinates come from four fixture AprilTags, recomputed every frame.

Thermal
MLX90640, 55° FOV

32 × 24 long-wave IR, −40 to 300 °C, 16 Hz. Chosen over the AMG8833, which saturates at 80 °C and cannot see solder temperatures at all.

Actuation
6 × Feetech STS3215

30 kg·cm serial-bus servos with 4096-count magnetic encoders, daisy-chained on one wire and one 12 V rail. One drives the rack-and-pinion carriage along the arch.

Feeder
Pancake NEMA17

A silicone-faced drive wheel against a sprung idler — compliant grip instead of a knurled gear, so soft flux-cored wire isn't crushed.

Trigger
27 mm piezo disc

Sandwiched under the fixture plate. Two impulses 80–500 ms apart, gated on the camera seeing the tip near the board.

Travel
180° arch rail

Six F625ZZ flanged rollers carry the carriage foot-to-foot over the apex. Carbon-filled PETG at the carriage, rack, and brackets; standard PETG for the links.

Project Ferros arm and sensor arch on the workbench images/ferros-hero.jpg
2400 × 1350 px · 16:9
Fig. 01 — Assembled arm under the sensor arch

The build

Every part, under $300.

— priced at the stall, not the catalogue

The whole point is that this is reproducible. Sourced from the Huaqiangbei markets in Shenzhen, with a walk-away price on every line so the budget survives contact with a stall.

Budget cap
$300

Excludes tools and compute already on the bench.

Parts total
~$232

Roughly $68 of headroom held for spares and a backup thermal module.

Encoded axes
6

Every joint holds position on its own, on a single serial bus.

PartSpecQtyCost
Motion — encoder servo arm
Feetech STS3215, 12 V30 kg·cm, 4096-count magnetic encoder, serial bus — all six axes6$70.85
STS3215 spareIdentical part, interchangeable with any axis1$11.80
TTL half-duplex adapterFeather S2 UART to the Feetech single-wire bus1$1.40
Pinion gear, metal5 mm bore, meshes the printed shoulder-pitch arc rack1$1.10
Turntable bearingSteel, 80–120 mm, load-rated for base yaw1$3.90
F625ZZ flanged bearing5 mm bore — arc-rail carriage rollers6$1.65
625ZZ / 688ZZ bearingElbow and wrist hinge shaft support8$4.45
Steel shaft rod5 mm and 8 mm diameter, cut to length1 lot$4.85
Fastener assortmentM3 / M4 / M5 bolts, nuts, washers, standoffs1$4.45
ThreadlockerBlue, medium strength1$1.10
Subtotal$105.60
Solder feed — cassette feeder
Pancake NEMA1720–25 mm body — kept as a stepper for step-counted linear metering1$2.20
TMC2209 driverUART, 2 A, silent — StallGuard doubles as jam detection1$1.95
Shaft coupler / hub5 mm bore, set-screw — mounts the printed drive wheel1$1.40
Silicone O-ringsStretched over the drive and idler wheels as a soft grip surface1 pack$1.40
Compression springLight rate, 10–15 mm — idler tensioner1$0.70
625ZZ bearingIdler and drive-wheel support2$1.10
PTFE guide tube1.0–1.2 mm ID — tight fit is what stops the wire buckling2 m$1.40
Blunt needle kitStainless 14–18 ga, Luer-lock — the last 25 mm of the path1 kit$2.50
608ZZ bearing8 mm bore — spool spins free in the quick-load bay2$0.55
Subtotal$13.20
Perception & trigger
MLX90640 breakout55° FOV, 32 × 24, −40 to 300 °C, 16 Hz — verified module, not a parts-bin clone1$29.20
Piezo disc27 mm — tap trigger under the fixture plate3$0.20
K-type + MAX31855Thermocouple ground truth for the melt-signature calibration1$4.20
AprilTag sheet36h11 family — fixture plate and end effector1$1.10
Subtotal$34.70
Electronics & power
12 V 5 A supplyServo bus draws 1.5–2.5 A realistic, ~4 A worst-case transient1$3.90
Buck converter12 V to 5 V, 3 A — logic rail1$0.70
E-stop, NC mushroom22 mm panel mount — breaks the 12 V rail in hardware, not in software1$1.40
Wire and shielded pairServo bus, piezo run, power — the iron is the loudest noise source on the bench1 kit$3.05
PerfboardHand-soldered carrier circuit2$1.10
Subtotal$10.15
Filament, fixture & consumables
Bambu PETG-CF1 kg spool — ~430 g into the joints, carriage, rack, and brackets1$32.00
Bambu PETG1 kg spool — links and non-load-bearing brackets1$22.00
Fume extraction120 mm 12 V fan and carbon pad — interlocked, no extraction means no feed1$2.75
Solder wire, SAC3050.8 mm (~100 g) and 0.6 mm (~50 g), lead-free by choice2$3.90
Flux pen, brass woolBrass wool, not a wet sponge1 ea$2.10
Mat, ties, sleevingESD-safe silicone mat, zip ties, heatshrink1 lot$5.15
Subtotal$67.90
Total against a $300 cap~$232

Prices are goal figures at roughly 7.2 RMB to the dollar; each line also carries a walk-away maximum, and the first quote at a stall routinely runs 30–80% high. Filament is ordered online rather than bought at a stall.

Already on the bench — not in the budget

Raspberry Pi 5 OAK-D PoE + injector Feather S2 3D printer Soldering station Hand tools

Where it stands

Build log.

  • DONE

    Engineering package

    An 18-page write-up covering architecture, the sensor trade study, feed mechanics, safety and risk analysis, a phased build plan with exit criteria, and a test plan built for ARKSEF.

  • DONE

    Sourcing and BOM

    Full parts list costed against a $300 cap, with a per-line walk-away price and a market-by-market sourcing map.

  • PHASE 1 · NEXT

    Feeder bench prototype

    Characterise commanded versus actual feed length, slip against idler tension, and jam modes. Exit criterion: ±0.5 mm over a 10 mm feed, 50 consecutive cycles jam-free.

  • PHASE 2

    Melt detection

    The emissivity study across SAC305 and Sn63, against thermocouple ground truth. Exit criterion: fire within ±0.4 s of confirmed liquidus on at least 18 of 20 trials, with zero false fires on 20 heat-without-melt trials.

  • PHASE 4

    Ferros Zero

    The full sensing chain with the nozzle positioned by hand — tap designates, thermal gates, feeder feeds and retracts. Deliberately a complete result on its own, so nothing downstream is load-bearing on the arm being finished.

  • FERROS 2

    Roadmap

    A second arm carrying the iron itself, turning the bench into a fully automated point-soldering cell at roughly a third the price of the cheapest machine that removes the human trigger.

Ferros is in the planning-to-build transition. Photos, source, and measured results go up here as each phase lands.