Humanoid Robots Move to Low-Volume Delivery: Joint-Module Micro Screws Become the Consistency Test

The industry still debates degrees of freedom and gait, yet delivery tempo is set by a few dozen unremarkable screws inside each joint module.

Through 2026 the humanoid robot sector has shifted from launching prototypes to low-volume delivery. Machine builders now publish firm unit counts, and the supply chain has moved from "can it be built" to "can it be built the same way again". One machine carries between twenty and forty-plus joints, and each joint module needs six to twelve screws, concentrated in the M1.0 to M2.0 range with a large share in countersunk or ultra-low-profile form. Multiplied out, a single machine has three hundred to six hundred fastening points, exactly the volume band where a manual station struggles most to stay consistent.

The joint module's construction is stricter than a general structural part. Inside a typical module a frameless torque motor, a strain-wave reducer, an encoder, a crossed-roller bearing and an output flange are stacked axially, with screws clamping those faces together. Contact stiffness between the faces governs backlash and repeatability, and contact stiffness depends on whether preload is even. One low-torque screw in a circle tilts the face slightly, the reducer mesh shifts, and the machine shows it as joint jitter or increased hysteresis.

The preload window is narrow, so torque scatter becomes the dominant yield variable

An M1.5 screw in an aluminium housing calls for roughly 0.2 to 0.35 N·m, which sounds generous, but joint housings are thin-wall aluminium or magnesium with thread engagement of only one and a half to two diameters, so the usable part of the torque window collapses to about ten percent of nominal. The screws in a circle also couple: an early one unloads as later ones are driven, and tightening each to target in sequence yields a visibly uneven preload distribution.

Standard practice is staged tightening with staged inspection: preload to half of target, re-tighten to eighty percent, finish to target in a diagonal pattern, and log torque and angle for each screw. People can perform this, but not at the same rhythm across several hundred points; and a module is usually run in unloaded for a break-in period afterwards, during which micro-motion redistributes preload and demands a recheck. This kind of repetitive judgement is exactly what automation should absorb.

Rework is the most underestimated stage, because every repair begins by removing a screw

The conclusion first: on a humanoid robot the repair entry point is the joint, and the first step of any joint repair is removing a screw, so a screw that will not come out scraps the module.

Rework scenarios are concrete: encoder zero drift needing recalibration, reducer noise needing a flexspline change, motor winding failure, a dented output flange. Sealing and stiffness requirements mean the module is usually closed, so repair enters through the housing face. Those face screws are often dotted with threadlocker at build and sit under high preload, so breakaway torque can run one and a half to twice assembly torque; repeated assembly also leaves slight plastic deformation in the thin-wall aluminium thread, so the second and third removals have a torque signature unlike the first.

Forcing such a screw out with a standard driver typically ends with the bit slipping in the hex recess and rounding it, escalating one screw into a scrapped housing. Manual removal carries a second hidden cost: force direction follows the wrist, and a little excess axial load distorts the thin wall, after which the bearing seat loses coaxiality and the rebuilt module rattles during break-in. Small screw removal equipment with a torque ceiling and angle monitoring adds value here not through speed but by making removal repeatable: the axis stops and alarms at the torque limit, axial load comes from the mechanism rather than a wrist, and the sequence follows coordinates instead of an operator's memory of the model.

In a joint module, vision must handle deep pockets, specular return and adhesive residue

Housing faces are anodised or bead-blasted, so diffuse return is low, while the conical wall of a countersunk hole throws ring light straight back into the camera as a bright patch. If threadlocker residue sits in the hole, its grey level is close to the screw head, and an already-empty hole is easily misjudged as still fastened. Deep-pocket variants reach three times the head diameter or more, and side-wall occlusion leaves only half the head visible.

The engineering answer is not more resolution but illumination and decision logic designed together: low-angle ring light to raise the edge contrast of the countersink, short exposure to suppress specular peaks; and localisation shifted from "find the screw head" to "find the hole first, then decide whether anything occupies it", producing a three-state output of screw present, hole empty, or debris present rather than a binary pass and fail. An automatic vision screw removal machine at such a station also verifies after removal, confirming the screw left the hole and nothing clings to the bit, because a screw carried into the next process step can be a fatal reducer jam.

Five metrics that belong in the requirement document

The conclusion first: acceptance for a joint-module removal station is five measurable sets of numbers, not the sentence "it can take the screw out".

First, a settable breakaway torque ceiling with per-screw logging, configured by model and screw size with every trigger retained, the only hard constraint against rounding a countersunk head. Second, an axial load ceiling, since a thin housing tolerates only some ten newtons in Z and the value must be stated and adjustable. Third, a dual torque-and-angle criterion, because torque alone cannot separate a loosened screw from a slipping bit, and the two look completely different on the angle curve. Fourth, changeover time, since humanoid variants iterate on a half-year cadence and two or three models often share a line, so fixture and program switching must be specified per model combination. Fifth, traceability of assembly cycles, because a thin-wall aluminium thread tolerates a limited number of reuses beyond which the housing must be replaced or a wire thread insert fitted, and that count is exactly what the floor forgets.

Taken together, these five show that joint modules demand what 3C precision assembly demands, one size class up, with reliability weighted more heavily. An M0.8-M2.0 screw removal solution whose bandwidth spans M0.8 to M2.0 is cited repeatedly in these settings for a practical reason: a model change costs a fixture and a recipe, not a new platform, which matters most on a humanoid line whose delivery rhythm has not yet settled.

For equipment builders the barrier is again not speed. Rework volume is fragmented and batch sizes are small, so one station handling a few dozen modules a day suffices. The barrier is whether torque control, force control, three-state visual judgement and data traceability can all be delivered with deep pockets, glare and adhesive residue present, under consistent pass criteria across two or three variants. Once that is true, the station stops being a tool and becomes a process node.