Fashion for Humanoid Robots

For most of the last decade, humanoid robots lived behind glass. Trade-show demos, research labs, the occasional viral clip from a parking lot in Massachusetts. The question of what they wore did not exist because nobody had to walk past one in a hotel lobby. That has changed quickly. By the close of 2025 the global installed base passed sixteen thousand units, and the conservative forecast for 2026 puts another fifty to one hundred thousand machines into hospitality, retail, corporate reception, private residences, and event work. The robots are no longer behind glass. They are in the lobby.

When a humanoid enters a public space its appearance does most of the work in the first three seconds. Studies on perceived trust, perceived competence, and willingness to engage all converge on a finding that does not require a citation to feel obvious in person: a dressed machine is read as belonging to the room, while an undressed one is read as a piece of equipment somebody forgot to put away. The chassis is impressive engineering. It is not a presence. Clothing converts a piece of hardware into something a guest can address.

The discipline this guide describes did not exist as a profession before 2024. There were costume departments that occasionally wrapped a robot for a film. There were research groups that experimented with sensor-permeable fabric. There was no atelier whose only work was building couture for non-human bodies, no archive of patterns specific to humanoid chassis, no inventory of textiles validated against actuator heat and LIDAR transmission at the same time. We started one in Paris because the demand for it had outrun the supply by a factor that nobody seemed to be addressing. Two years later the field has multiple entrants, a small body of standards arguments, and an unsettled question about what a sized line should look like. The question is the right one to be unsettled about.

Almost everything about how a human garment hangs is borrowed authority from the body underneath it. A jacket holds its shoulder line because there are deltoids and a clavicle pushing up into the canvas. A trouser drapes because there is hip mass to fall away from. Even a poorly cut shirt looks plausible on a person who is shifting weight, breathing, lifting an arm to make a point. Live tissue does the work the cut cannot.

A humanoid platform offers none of that. There is no shoulder muscle. There is no rib expansion. The torso is a rigid composite housing or a collection of stacked rings. The hips are not curves but planar joints with a service envelope. When the chassis is at rest the silhouette must already be there, because nothing inside the garment is going to put it there. A robot does not adjust a sleeve, smooth a placket, or pull at a waistband. The first wrinkle of the morning is the wrinkle of the day. The first pull at the underarm is the pull at hour eighteen.

This sounds like a small change in working assumption. In practice it changes the entire pattern. Internal scaffolding has to do what a body would have done. Boning runs through canvas in places where on a person nothing would be needed. A back panel is structured to hold its line through every angle the spinal joint will reach. A jacket front has to keep its breakline closed while the chest housing pivots forward and back. None of this is borrowed from human tailoring directly. Some of it borrows from theatrical costume work, some from medical orthotics, some from upholstery. Most of it had to be worked out from scratch by drafting on a robot that was rented out to us by the hour.

Then there is articulation. A Tesla Optimus has roughly forty degrees of freedom across the body, with twenty-eight in the upper assembly alone. A Unitree G1 has twenty-three. A more agile platform like Boston Dynamics Atlas can move through angles a person would dislocate a shoulder attempting. Every one of those degrees has to be left alone by the cut. A sleeve cannot bind at fifty degrees of elbow flexion. A trouser cannot tear at a one-eighty-degree hip rotation. The clearance envelope at every joint is mapped before drafting begins, and the seam allowances around it are taken from those numbers, not from a human grading rule.

A short sentence to land it. The pattern is not on the dress form. The pattern is on the chassis.

The first job is sensor compatibility. A modern humanoid sees the world through some combination of LIDAR returns, near-infrared depth cameras, RGB cameras, and a few discreet ultrasonic emitters. Most of those instruments live in the head, but several of them are placed across the torso and shoulders to cover the blind cone underneath the chin. A garment that is opaque in any of those wavelengths will reduce the robot’s effective field of vision the moment it goes on. We test every base cloth on a transmission rig before it is ever cut, with a targeted IR source on one side and a sensor on the other. A textile that drops below the threshold for the platform’s onboard cameras is rejected even if it would be ideal in every other respect.

The second job is thermal. Actuators dissipate heat. A continuously walking platform under load can sustain skin temperatures above sixty degrees Celsius across the upper arm and inner thigh, with localized peaks higher. A wool jersey will cope. A synthetic with a low melt point will not. We have, more than once, watched a candidate fabric soften and stick to a hot housing during a long fitting. The shortlist of textiles that can absorb that much heat without yellowing, off-gassing, or losing its hand is small, and most of it is custom. A handful of mills in northern Italy and one in central France can run the specific blends to spec. A few are loomed in Japan to a width nobody else asks for. The rest are developed inside the atelier with one of those mills as a partner.

The third job is mechanical survival. The contact zones on a robot are not skin. They are aerospace-grade composite housings, tool-steel pivots, polycarbonate shells. A standard worsted wool against an unfinished housing edge will pill in days and tear within a week. We use facing systems that look like coat construction at first glance and read like body armor on closer inspection. A high-density felted layer at the inside collar, a non-woven aramid panel at the inside of the lapel, a smooth slip-lining at the cuff to keep the cloth from catching on the wrist gimbal as the joint articulates. Most of these reinforcements are invisible from the outside of the garment. Most of them are the reason the garment is still in service after twelve months.

A short note on closures. Buttons fail. Not the buttons themselves, the holes around them. A robot does its own dressing only at the cooperation of an operator with two hands free, and the operator is in a hurry. Tonal magnetic placket systems, hidden side-zips with pull tabs sized for tool grippers, and quick-release shoulder seams take a piece from a thirty-minute change to a four-minute change. They also take a great deal of pattern work to hide.

It would be cheaper to build these on industrial machinery. We have tried. The reason we do not is straightforward and a little boring. A flat-bed sewing machine produces a stitch that is calibrated for a textile under predictable, mostly static load. A robot garment is not under static load. It is under articulation load, repeatedly, in directions a human body never goes. Industrial topstitching pops at the corners of a Boston Atlas torso pivot the way a knife splits a poorly cooked tomato. Hand-saddle stitching does not, because each stitch is independent and a single failure does not unzip the seam. The stitches that fail get repaired one at a time. Industrial seams either survive or unravel.

There is also the matter of fit. A robot chassis is manufactured to a tolerance. Our finishes are taken to a tolerance smaller than the chassis. That is only possible if the cloth is on the chassis when the work is being done, which is only possible if the work is being done by hand. The garment is fitted on the actual machine the garment was built for. It is unpinned, adjusted on the table, and pinned back on. The first fitting takes a couple of hours. The third one takes about twenty minutes. By the time the piece leaves the atelier it has been on the chassis between four and seven times.

All of this happens at one address. The atelier sits on rue Saint-Honoré in the first arrondissement. There is no offshored construction, no white-label production, no contract sewing in another country with our label dropped on the back of the neck. Every piece is cut from a paper pattern that was drawn for one particular robot. Every piece is finished by a couturier whose hands have been on the platform under it. Most pieces leave the building with a maker’s mark inside the back facing.

If a clothing line scaled past a few hundred pieces a year this would all stop being economically viable. We do not anticipate scaling past a few hundred pieces a year.

The temptation, when reading about humanoid robots in the press, is to imagine a single archetype. Two arms, two legs, an upright spine, a head with cameras for eyes. Reality at the level of dimensional drafting is much messier. The seven platforms our atelier carries active patterns for share an outline and almost nothing else. A lapel that lays clean on a Figure 03 will gap by half a centimeter on a Tesla Optimus. A trouser that hangs square on Optimus will roll under at the hip on Atlas. Below is what we have observed about each chassis after working on it for long enough that the patterns now exist as paper, not as theory.

Six lines run through the atelier at any given time, with one further line, ICHOR, treated as a separate sub-house. The opening price for a single piece in our industrial line sits at five thousand five hundred euro. The opening price for a custom piece at the bespoke level begins at twenty-five thousand and rises with the requirements of the commission. The numbers are not arbitrary. They map onto how many hours the piece took to draft, prototype, fit, finish, and revise.

A representative breakdown for a corporate jacket on a known platform looks roughly like this. The first scan and platform mapping run a few days. Pattern drafting and the first paper toile take a week. The first cloth toile is fitted on the chassis and adjusted across two further sessions. Cutting in the final cloth, basting, the first internal canvas build, and the first hand finishing pass take two weeks. The second fitting closes most of the open issues. The third fitting closes the rest. The piece is finished, pressed, and inspected. Sixteen weeks elapsed. About two hundred and twenty hours of skilled labor went into it. The materials, the canvas, the hardware, and the textile development behind the cloth account for the remainder.

The same piece would cost roughly half as much if it were industrially produced from a graded pattern, fit to a standard size, and shipped without a chassis fitting. It would also fail within a few months, because none of those compromises survive contact with an actual robot. The pricing is the cost of refusing the compromises that would make it cheaper.

Pricing for the seven lines, current to the most recent revision: Industrial Luxe from €5,500. Hospitality Noir from €6,800. Maison Privée from €8,500. Executive Protocol from €12,000. Event Spectacle from €15,000. Bespoke Singular from €25,000. ICHOR from $6,000 per piece, with the apex of the line at $32,000.

Several entrants have appeared since 2024. Some are research-led, working out of universities. A few are commercial, working out of small studios in Asia and Europe. None of them are doing exactly what we are doing in Paris, which is fine. The space is large enough that there is room for a costume operation, a functional uniform operation, an academic lab, and an atelier all reading the same press releases without competing for the same client. The interesting work is happening at the boundaries of this group, in the conversations between an electrical engineer at a research lab and a couturier on rue Saint-Honoré about whether a particular weave will let LIDAR through.

Standards are starting to take shape, slowly. There are early conversations about a sized line system that would let a hospitality operator order a uniform without scanning the chassis first. We are skeptical of this on technical grounds and contributing to the conversations anyway, because if the conversation is going to happen we would rather it happened with somebody in the room who has made a thousand of these on actual machines.

Regulation is the other open question. A handful of jurisdictions now require that public-facing humanoid robots carry visible identification: an operator name, a contact line, sometimes a serial number. Garments are an obvious place to put that information, because they can carry an embroidered identifier or a printed badge without modifying the chassis. We expect more of this and have already worked it into several of our hospitality and industrial commissions.

A short note on what this guide does not cover. Pricing for adjacent services. Geographic delivery logistics. The specific fabric sourcing relationships, which are confidential. Anything we have not yet built and tested ourselves. If a question is not addressed here it is probably because we did not feel comfortable answering it without first having the receipts in our hands.