07 · The body
Materials, manufacturing and cost
What a robot is made of, how each part gets made, what the whole thing costs, and why the same humanoid costs a third as much when it is built from Chinese parts.
Three things set what a robot costs: how many you make, where you make them, and how many precision joints you put in. A Unitree G1's parts add up to about ¥41,574 (US$5,900) estimate. A Tesla Optimus Gen 2 built from Chinese suppliers comes to US$46,000 to 55,000, and the same robot with no Chinese parts about US$131,000 to 133,000 estimate. The gap is the cluster, the volume and the joint. This chapter shows you each one.
The cost of a robot on one screen
The bill of materials is the full shopping list of parts with a price next to each. It excludes the factory, the test, the warranty and the profit, so the selling price sits well above it. Unitree sells the G1 base for ¥85,000 including tax against that ¥41,574 parts list, and Unitree's whole company ran a 60.4% gross margin in 2025 verified. If you remember one ratio from this chapter, remember that one: in China the price of a robot is about twice its parts, and the parts are mostly joints.
What robots are made of
A robot body is a frame that must be stiff and light, joints that must be hard and precise, and a skin that must be soft and cheap. Each wants a different material. The frame is mostly aluminium, the gears and shafts are steel, the covers are glass-filled nylon, and the newest humanoids wear foam and fabric.
| Material | Typical robot use | Stiffness E, GPa | Strength, MPa | Density, g/cm³ | Max °C | Raw US$/kg |
|---|---|---|---|---|---|---|
| Aluminium 6061-T6 | frames, brackets, machined housings | 69 | 276 | 2.70 | ~150 | 2.5 to 5 |
| Aluminium 7075-T6 | highly loaded limbs and links | 72 | 503 | 2.81 | ~120 | 4 to 8 |
| Magnesium AZ91 | die-cast housings where grams count | 45 | ~150 | 1.81 | ~120 | 3 to 5 |
| Carbon fibre (CFRP) | long light links, premium shells | 50 to 60 | 500 to 700 | 1.55 | 120 to 180 | fibre 15 to 30 |
| PA66-GF30 nylon | covers, cable guides, light brackets | ~9.5 | ~180 | 1.36 | 150 to 180 | 3 to 4.5 |
| PEEK | hot or wearing plastic parts, bushings | 3.7 | ~100 | 1.30 | ~250 | 75 to 100 |
| Alloy steel 4140 | gears, shafts, screw nuts, bearing races | 200 | 650 to 1,000 | 7.85 | ~450 | 1.2 to 2.5 |
| Titanium Ti-6Al-4V | rare; roller screws and aerospace parts | 114 | 880 to 950 | 4.43 | ~400 | 20 to 50 |
Property values and prices from the materials primer, 2024 to 2025 bulk pricing estimate. Commodity prices move 30% or more in a year, so read the prices as ratios.
Real
Stiffness (Young's modulus E) comes from the bond between atoms and is fixed for a base metal, while strength comes from alloying and heat treatment and varies tenfold. Steel and aluminium have almost the same stiffness per kilogram, about 26 GPa per g/cm³. So an aluminium link as stiff as a steel one weighs about the same; aluminium wins because a thick aluminium wall resists buckling and machines three times faster.
ELI5
How bendy a metal is depends on what it is made of, and you can't change that. How hard it is to snap depends on the recipe, and you can. Aluminium and steel are equally bendy for their weight, so robot makers pick aluminium for the body because it's easy to cut and makes fat, rigid tubes.
Three material facts matter on a factory floor. First, "aerospace-grade aluminium" in a brochure means aluminium alloy 6061 or 7075, the standard bar any machine shop buys; the real aerospace content is the paperwork. Second, magnesium alloy is a third lighter than aluminium and die-casts well, which is why it turns up in housings. Third, carbon fibre is the stiffest per gram but is laid up largely by hand, so it pays only where grams are worth dollars.
Skins are where humanoids now differ most. Figure 03 is covered in soft textiles over multi-density foam at the pinch points instead of hard machined shells verified. 1X's NEO uses a 3D lattice polymer structure under a machine-washable nylon knit suit and weighs 30 kg verified. Xpeng's IRON wears a continuous soft skin with no visible joints verified. Soft skins are a safety choice for robots that work beside people, and they turn the cover from a moulded part into a garment that wears out and gets replaced.
Nylon datasheets quote strength "dry as moulded". PA66 absorbs 2 to 8% of its weight in water from the air over weeks, which roughly halves its yield strength and can drop its stiffness by 60 to 70% estimate. A glass-filled engineering plastic bracket that passes testing in an air-conditioned Shenzhen lab can creep and loosen in a humid Jakarta kitchen. The test condition matters as much as the number.
Money follows material into the joints. Unitree spent ¥794 million on raw materials in 2025, and ¥403 million of that (50.8%) was mechanical parts, mostly custom machined and die-cast pieces verified. Gear steel, bearing steel and screw alloys carry the precision. About 70% of planetary roller screws use titanium or chrome-moly steels, only about five foundries make the aerospace-grade alloy, and precision steel lead times run 6 to 9 months unverified. That is one reason the planetary roller screw is the slowest part to scale (see Gears and screws).
How parts get made
Five processes make almost every part in a robot. Two of them, CNC machining and 3D printing, need no custom tooling: you send a file and pay for every part. The other three, injection moulding, die casting and stamped sheet metal, need a steel mould or die first, and then each part costs cents to a few dollars.
| Process | How it works | Tooling | Per-part cost | Tolerance | Pays off from |
|---|---|---|---|---|---|
| CNC machining | a spinning cutter removes metal from a block | ~US$0 (setup ~US$250) | bracket: US$45 at 1, US$9 at 100, floor ~US$7 | ±0.025 mm routine, ±0.005 serious | 1 to a few thousand |
| 3D printing (nylon powder) | a machine fuses powder layer by layer | US$0 | bracket: US$12 falling to ~US$6 | looser than CNC | 1 to hundreds |
| Injection moulding | molten plastic is shot into a cooled steel mould | US$9k (1 cavity, China) to US$45k (4 cavities) | bracket: US$0.45, then US$0.22 | ±0.1 routine, ±0.05 good | ~1,500 parts |
| High-pressure die casting | molten aluminium or magnesium is forced into a steel die, 20 to 120 s a shot | US$20k to 80k; US$100k to 300k+ large | near-net shape, drill and face only | ±0.05 to 0.3 | ~10,000 parts |
| Sheet metal | laser cut and folded on a brake; progressive dies at volume | ~US$0; dies US$20k to 150k in China | folded by hand, then material plus pennies with a die | ±0.1 to 0.4 folded, ±0.05 to 0.2 die | any; dies from 50,000 |
Numbers for a palm-sized 50 g bracket and general process ranges, from the materials primer (2024 to 2025 service-bureau and China pricing) estimate.
Choosing between them is one sum. Total cost is the tool plus the marginal cost times the number of parts. Per part, that is the tool divided by the volume, plus the marginal cost. A tooled process starts absurdly expensive and falls toward its floor; a tool-free process stays flat. The volume where the lines cross is the whole decision.
Real
Cost per part = T ÷ n + c, where T is the tool and c the marginal cost (material, machine time, labour, scrap). Break-even between a tool-free process 1 and a tooled process 2 is n* = T₂ ÷ (c₁ − c₂). Because c₂ for moulding is cents, a printer getting 50% cheaper moves n* by less than one order of magnitude. Where the tool is built matters more: a Shenzhen tool at a quarter of the Western price moves the crossover about 4x.
ELI5
Taxi or car. A taxi (printing, machining) costs the same every ride. A car (a mould) costs a fortune up front, then each ride is nearly free. For ten rides take taxis; for ten thousand buy the car. Shenzhen sells the cheapest cars in the world, so there the mould pays off sooner.
Robots live on the left of this chart for longer than phones do. Figure had delivered about 350 Figure 03 units by April 2026 and Unitree shipped about 5,500 humanoids in 2025 verified. At those volumes a humanoid is mostly machined and die-cast metal plus a few moulded covers, which is what Unitree's purchase ledger shows. Every product with 10 to 20 custom parts needs about US$100,000 to 500,000 of tooling before the first unit ships estimate, and that money is spent when you know least about what customers want.
Design for manufacture and tolerances
DFM (design for manufacture) and its twin DFA (design for assembly) are the rules that let a factory make your part cheaply and assemble it fast. The standard method, Boothroyd and Dewhurst's, asks three questions of every part: does it move relative to its neighbours, must it be a different material, and must it come apart for assembly or service? If the answer to all three is no, merge it into its neighbour. Fewer parts means fewer tools, fewer suppliers and fewer steps where a worker can get it wrong.
Precision is bought in steps. Every dimension on a drawing carries a tolerance, the wiggle room a part may have and still pass. For machined metal, ±0.1 mm is cheap; ±0.025 mm adds 30 to 100% to that feature; ±0.01 mm needs grinding and real measurement and costs 2 to 5 times as much; ±0.0025 mm needs lapping in a temperature-controlled room at 5 to 20 times estimate. One careless decimal place on a drawing picks which shops, and which country, can make your part.
Real
A process makes a spread of sizes. Cpk counts how many standard deviations fit inside your tolerance band: Cpk 1.0 is about 2,700 defective parts per million, 1.33 (the usual purchasing contract figure) about 63, and 1.67 about 0.6. The number that matters is the Cpk a supplier holds at a given tolerance and price, because anyone can hold a tolerance briefly.
ELI5
No two parts come out the same size. If the sizes wobble a little and your allowed window is wide, almost every part fits. If the window is tight, some fall out and get thrown away, and you pay for those too.
Joints are where the tight tolerances live. A harmonic drive's flexspline, wave generator and assembly and test make up more than 70% of its unit cost estimate, because the teeth only mesh if the thin steel cup is machined and ground to microns. The G1's small joint is a 4-in-1 joint module (motor, planetary reducer, encoder and driver) with a hollow shaft for wiring estimate; every one of those fits is a tolerance someone pays for.
"The last 10% is 90%." Getting from a prototype that works on your bench to units shipping at good yield takes as long as everything before it, and the time goes on unglamorous items: tolerance stacks, connector insertion, cosmetic defect limits. Ten prototypes tell you the average part. Only hundreds tell you the 2% that fail. Plan your first production run to find those, and price it that way.
Assembly lines, test and yield
A robot factory has four jobs: check incoming parts, build sub-assemblies such as joint modules, run final assembly, and prove each robot works before it leaves. The best public description of a humanoid production line is Figure's BotQ: more than 150 networked workstations on custom manufacturing software, more than 50 in-process inspection points, and hundreds of suppliers qualified against incoming inspection verified Apr 2026. Figure says it went from one robot a day to one an hour, 24 times faster, in under 120 days, on a first line rated at 12,000 robots a year.
Yield is the share of units that pass. Burn-in testing runs each finished robot for hours before shipping so that parts destined to fail early (a weak solder joint, a bad encoder cable) fail in the factory instead of at a customer. No humanoid maker publishes its burn-in duration or yield. For a reference point from consumer hardware, cosmetic anodised aluminium parts historically run 60 to 90% first-pass yield through the mould, machine, blast and anodise chain estimate.
Capacity claims are everywhere; output is rarer. Agility's RoboFab in Salem, Oregon is 70,000 sq ft rated at 10,000 robots a year verified. 1X says its Hayward, California factory can build 10,000 NEOs a year unverified. Xpeng announced a 110,000 m² humanoid production base in Guangzhou's Tianhe district in Feb 2026 verified. Leju builds its Kuavo-5 at a Foshan plant at about one every 30 minutes, about 10,000 a year unverified. Compare every one of those with the roughly 150 humanoids each that Figure, Agility and Tesla shipped in 2025 estimate.
The humanoid bill of materials
Across every estimate, actuators are the biggest line: 40 to 60% of a humanoid's parts cost (about 56% for Optimus), then sensors at 10 to 25%, structure at 10 to 20%, compute and control at 10 to 15%, and battery at 5 to 10% estimate. Hands alone are about 17% of an Optimus, roughly US$9,500 a pair. Bank of America's 2030 split makes the point sharper: linear actuators 27%, rotary actuators 24%, dexterous hands 19%, and everything else about 30% estimate.
The G1 teardown puts numbers on each line: 23 joint modules at ¥27,500 (14 small ones at ¥1,000, 9 large at ¥1,500), a Livox MID-360 lidar at ¥3,840, an Intel RealSense D435i depth camera at ¥1,869, a Rockchip RK3588 compute board with memory at ¥1,415 and ¥3,000 of processing and assembly estimate. Inside one small 60 mm joint, the planetary reducer is ¥300, the motor ¥200, the encoder ¥200 and the structure ¥150. SemiAnalysis modelled a different G1 configuration at US$8,976 contested. Read both as a range and the conclusion holds: joints are half to two thirds of the robot.
Retail prices show the gap between a volume buyer and you. Leju pays about US$202 per actuated axis (US$156 for the module, US$46 for the driver) at about 36,000 modules a year estimate. On the open market, a DAMIAO DM-J8009-2EC joint is US$385, a MyActuator RMD-X8-25 US$450 to 480 and a ZeroErr eRob 80T from US$1,440 verified Sep 2026. A small company buying ten joints pays two to seven times what a scale buyer pays.
The cost-down curve
Cost-down is the steady cut in unit cost as volume grows and suppliers compete. Its rule of thumb is the learning curve: every time cumulative volume doubles, unit cost falls by a roughly fixed percentage. Leju's joint gives you a clean measured case.
| What falls | From | To | Source |
|---|---|---|---|
| China-sourced humanoid parts | US$35k (2025) | US$13k to 17k by 2030 to 2035 | BofA estimate |
| Average humanoid selling price | US$114,700 (2024) | ~US$37,000 (2030) | IDTechEx estimate |
| Frameless torque motor | ¥2,000 to 5,000 | below ¥1,000 | gelonghui, eastmoney estimate |
| Coreless finger motor | ¥1,000 to 2,000 | ~¥600 | Moons' via stcn estimate |
| 6-axis force sensor | ¥10k to 20k | ¥1k class | Sohu market report estimate |
| Unitree average humanoid price | ¥590k (2023) | ¥168k (2025) | prospectus verified |
| NVIDIA Jetson T5000 module (1k units) | US$2,999 | US$4,999 (up) | CNX Software, 22 Jul 2026 verified |
One line in that table goes the wrong way. NVIDIA raised Jetson prices by up to 101% on 22 Jul 2026; the AGX Thor developer kit went from US$3,499 to US$5,499 verified. Observers blame memory prices. Compute is a 10 to 15% line in a humanoid today, and it is the one line a Chinese supplier cannot yet undercut for Western robots. Budget it rising.
Capacity is running ahead of demand, and that is what drives the cost-down. Leadshine plans capacity for 2 million frameless motors in 2026 after delivering 120,000 in 2025, and Chinese screw makers are building millions of roller-screw sets, against UBS's projection of about 30,000 humanoids shipped worldwide in 2026 estimate. That glut is good for buyers and bad for anyone whose plan is to make a commodity part.
Why China is cheaper
Wages explain the least. In Unitree's cost of goods, direct materials are 81.7% and direct labour 8.1% verified. Four other things explain the gap.
- Cluster density. A Shenzhen, Dongguan and Huizhou maker finds its suppliers within about 200 km estimate. Tooling comes back in days, a failed part is redesigned the same week, and every component class has a dozen competing vendors. The Shenzhen and Dongguan cluster roughly halved world tooling prices estimate.
- Car parts reuse. The electric-car supply chain already makes motors, magnets, screws, batteries and die castings at millions of units. Wuzhou Xinchun's planned roller-screw expansion is sized for about 80,000 humanoids plus 1.35 million cars estimate. Optimus's actuator Tier-1s, Sanhua and Tuopu, are car-parts companies, and Optimus V3 is about 70% Chinese content estimate.
- Engineering and machine time. A CNC machine and operator costs about US$12 to 40 an hour in China against US$40 to 150 in the West, and an injection press US$15 to 40 against US$40 to 120 estimate. A Chinese mould at about US$9,000 against a Western US$40,000 moves the break-even volume 4x.
- Vertical integration and speed. Unitree makes its own joint drives, mechanical structure and battery management, and designs its own planetary tooth profile to survive leg impacts; bought-in parts (cells, compute modules, memory, some sensors) are only 14 to 18% of its cost verified. SemiAnalysis estimates Unitree's in-house motors, gearboxes, lidar and cameras cost 30 to 40% of Western equivalents estimate.
Chinese component makers are winning inside China first. Harmonic Drive Systems, the Japanese inventor of the part, held only 4.4% of China's harmonic reducer market in 2025, against Leaderdrive at 27.5% and Laifual at 21.4% estimate. Domestic makers hold 92.8% of China's humanoid 6-axis force sensor market estimate. When a Shenzhen supplier says its part is a domestic substitute for a Japanese one, that is now the normal case.
OEM, ODM, contract manufacturing and the deal
You can build a robot three ways without owning a factory, and the words for them decide who owns the design and the tools.
| Model | Who designs | Who owns the design | What you pay for | Robotics example |
|---|---|---|---|---|
| OEM contract build | you | you | parts, labour, NRE, tooling | Jabil building Apptronik's Apollo verified |
| ODM | the factory | usually the factory | a finished design, often a minimum order | a maker's standard robot re-badged |
| EMS | you | you | electronics assembly and test only | Foxconn, Luxshare, BYD Electronics unverified |
| White label | the maker | the maker | finished units under your brand | common for service robots |
Contract manufacturing is a thin business. Foxconn's net margin has hovered around 2 to 3% for a decade, and toolmakers earn 5 to 15% estimate. That thinness shapes every negotiation: the factory makes its money on volume and on your mistakes, so the contract terms matter more than the quote.
- Sample, prototype, small batch, mass production. A sample is one part to check; a prototype is a working unit; a small batch (tens to hundreds) proves the line; mass production is when the tools are hard and the price drops. Each step has its own price.
- Tooling ownership. Put in writing that you own the moulds you paid for and can move them. Toolmakers usually take about 50% up front estimate.
- MOQ. The MOQ (minimum order quantity) is the smallest batch a supplier will run. It is negotiable against a credible forecast.
- Quotes are opening bids. Tooling and piece prices flex 20 to 50% against volume commitments and competitive tension, and the first quote prices in how messy your drawing looked estimate.
- Cash float. Hardware pays out months before it gets paid: deposits, parts on 30-day terms, six weeks at sea, customers on 60 to 90 days. That is a 90 to 180 day gap on every unit estimate, so growing faster burns cash faster.
Certifications to sell in Southeast Asia
Certification for robots today is type testing before sale: a lab tests one design against a standard and you get a mark. Singapore's National Robotics Programme lists the standards that apply there, including SS 713:2025 for robots talking to lifts and doors, SS ISO 13482 for personal care robots and SS ISO 10218 for industrial robot safety verified. TÜV SÜD PSB tests robot safety in Singapore, and SGS tests robots in Singapore, Malaysia and the Philippines and issues ISO 13482 certificates; the labs test electrical, mechanical, functional safety, cybersecurity, EMC and performance verified. No recurring in-service inspection rule for service robots turned up in Singapore unverified.
| Mark | What it covers, in plain words | Where |
|---|---|---|
| CE | the maker declares the product meets EU safety, EMC and radio rules; often accepted as evidence in Southeast Asia | EU |
| FCC | radio emissions and interference from anything with Wi-Fi, Bluetooth or a fast processor | US |
| UL | electrical and fire safety, including battery packs, tested by an independent lab | US, used as a benchmark widely |
| IMDA | registration of radio and telecom equipment such as Wi-Fi and 4G modules | Singapore |
| SIRIM | Malaysia's certification body, including radio equipment approval | Malaysia |
| SDPPI | Indonesia's type approval for radio and telecom devices | Indonesia |
| RoHS | limits on lead, mercury and other hazardous substances in electronics | EU origin, copied widely |
| UN38.3 | transport tests a lithium battery must pass before it can ship by air or sea | worldwide |
Scope lines are plain summaries; the exact rules for robots in each Southeast Asian country were not verified for this primer unverified. Confirm with a lab before you import.
What makers already hold shows the usual set. Gausium holds ISO 9001, CE, UL and TÜV certificates verified. Figure's Figure 03 battery is UN38.3 certified and targeting UL verified. AgiBot says its A2 is certified for China, the US and the EU unverified. A robot with a radio inside needs the local radio approval in each country you sell into, and one with a battery needs UN38.3 before it can be shipped at all; see Compute, wiring and power for the battery shipping rules.
Terms in this chapter
| English | 中文 + pinyin | ELI5 | Real definition |
|---|---|---|---|
| Bill of materials (BoM) | 物料清单wùliào qīngdān | The full shopping list of parts in one robot with the price of each. Add it up and you know what the parts cost before the factory, the test, the warranty and the profit. | The structured list of every component, sub-assembly and quantity in a product with unit costs. A Unitree G1 base is about ¥41,574 (23 joint modules ¥27,500, lidar ¥3,840, depth camera ¥1,869, compute ¥1,415, assembly ¥3,000; China Post Securities, Mar 2026); an Optimus Gen 2 from Chinese suppliers US$46k to 55k (Morgan Stanley). Actuators are 40 to 60% of a humanoid BoM. |
| Cost-down | 降本jiàngběn | Making the same part cheaper every year, by buying more, designing it simpler and making suppliers compete. | The planned reduction in unit cost through volume, redesign, supplier competition and localisation. Leju's joint module plus driver fell from US$433 to US$202 per axis over three years as purchases grew 100x (corematter, Sep 2026). |
| Tooling | 开模 / 模具费kāimú / mújù fèi | The custom steel moulds and dies a factory has to make before it can stamp out your part. Expensive once, then every part is cheap. | Design-specific moulds, dies, jigs and fixtures, paid up front and usable for one part design only. A single-cavity Chinese injection mould is about US$9,000; die-casting dies US$20k to 300k+. A product with 10 to 20 custom parts needs about US$100k to 500k of tooling (materials primer, estimate). |
| Mould | 模具mújù | A steel block with a part-shaped hole inside. Squirt in melted plastic or metal, let it cool, pop the part out, repeat. | The hardened steel tool that forms a moulded or cast part, with its cavity, cooling channels and ejectors. Die-casting dies in H13 steel last about 50k to 150k aluminium shots. Contracts should state who owns the mould and whether you can move it. |
| Die casting | 压铸yāzhù | Squirting molten aluminium or magnesium into a steel mould at high pressure, like a waffle iron for metal. | High-pressure die casting (HPDC): molten aluminium, zinc or magnesium injected into a steel die at 10 to 100+ MPa; 20 to 120 s cycles, ±0.05 to 0.3 mm, tooling US$20k to 300k+, economic from about 10,000 parts. Parts are porous, so usually not welded or fully heat treated. |
| Injection moulding | 注塑zhùsù | Melting plastic pellets and shooting them into a cold steel mould. Every LEGO brick is made this way. | A screw melts polymer and injects it at 500 to 2,000 bar into a cooled mould; cycle time is set by cooling, about 2 to 4 s per mm² of wall. A US$9,000 single-cavity tool makes a 50 g bracket at about US$0.45 and beats printing above about 1,500 parts. |
| CNC machining | CNC加工 / 机加工CNC jiāgōng / jījiāgōng | A computer-driven cutter carves a part out of a solid block. No mould needed, but you pay for every minute of cutting on every part. | Computer-controlled milling and turning from bar or plate. No tooling; ±0.025 mm routine, ±0.005 mm in a controlled shop. Machine plus operator about US$12 to 40 an hour in China, US$40 to 150 in the West (estimate). The default for joint housings, reducer parts and low-volume robot structure. |
| Sheet metal | 钣金bǎnjīn | Cutting flat shapes out of thin metal sheets and folding them into boxes and brackets, like metal origami. | Laser or punch cutting plus press-brake bending (no tooling, ±0.1 to 0.4 mm) at low volume; progressive dies (US$20k to 150k in China) at high volume. Used for enclosures, battery trays, chassis and brackets. |
| Aluminium alloy | 铝合金lǚhéjīn | Light, stiff, cheap metal that is easy to cut. Most robot frames are made of it. | 6061-T6 (E 69 GPa, yield 276 MPa, 2.70 g/cm³) for general structure and 7075-T6 (yield 503 MPa) for highly loaded links; raw US$2.5 to 8 per kg (estimate). "Aerospace-grade" in marketing usually means one of these two. |
| Magnesium alloy | 镁合金měihéjīn | A metal a third lighter than aluminium that pours well into moulds, used for light housings. | AZ91 and similar: E 45 GPa, about 150 MPa, 1.81 g/cm³, service to about 120 °C; die-cast for housings where grams matter. Needs corrosion protection. |
| Carbon fibre | 碳纤维tànxiānwéi | Super-stiff black threads glued in resin. Very light and strong, but slow and costly to make. | Carbon fibre reinforced polymer (CFRP): quasi-isotropic laminates E 50 to 60 GPa at 1.55 g/cm³; fibre US$15 to 30 per kg, prepreg US$50 to 150 (estimate). Largely hand laid up, so it pays only where weight is worth money. |
| Engineering plastic | 工程塑料gōngchéng sùliào | Tough plastics that can replace metal in covers and brackets, like glass-filled nylon. | Polymers such as PA66-GF30 (about 9.5 GPa, 180 MPa, 150 to 180 °C), POM and PEEK (to about 250 °C, US$75 to 100 per kg). Nylon absorbs 2 to 8% water, which can halve its strength; check test conditions. |
| Tolerance | 公差gōngchā | How far a part is allowed to be off its drawing size and still pass. Tighter costs more, a lot more. | The permitted variation of a dimension. Machined metal: ±0.1 mm cheap, ±0.025 mm +30 to 100%, ±0.01 mm 2 to 5x, ±0.0025 mm 5 to 20x (estimate). Suppliers are judged by Cpk at a tolerance; 1.33 is about 63 defects per million. |
| Assembly | 装配zhuāngpèi | Putting all the parts together into a working robot, by hand, by machine or both. | Joining sub-assemblies (joint modules, hands, torso, wiring harness, shell) into a finished unit. Design for assembly cuts part count and steps; Unitree's processing and assembly is about ¥3,000 of a G1's ¥41,574 BoM (China Post Securities, estimate). |
| Production line | 产线chǎnxiàn | The row of workstations a robot moves through as it gets built and checked. | The sequenced stations, fixtures, software and inspection points that build a product. Figure's BotQ: 150+ networked workstations, 50+ in-process inspection points, one robot an hour by Apr 2026 on a line rated at 12,000 a year. |
| Yield | 良率liánglǜ | Out of every 100 robots you build, how many pass the checks first time. | The share of units passing inspection. First-pass (rolled throughput) yield is the product of station yields: three stations at 95% give about 86%. Cosmetic anodised parts historically run 60 to 90% first-pass (estimate). Humanoid makers do not publish yields. |
| Burn-in test | 老化测试lǎohuà cèshì | Running each new robot for hours before it ships, so the weak ones break in the factory instead of at the customer. | Powered operation of finished units under load or heat to trigger early-life failures (the left of the bathtub curve) before shipment. Durations for humanoids are not public; ask each maker. |
| Contract manufacturing | 代工dàigōng | Paying another company's factory to build your product for you. | Outsourced production by a firm that owns the factory but not the product. Margins are thin: Foxconn's net margin has hovered around 2 to 3% for a decade. Jabil builds Apptronik's Apollo; Foxconn led the private placement in Agility's SPAC. |
| ODM | ODM(原始设计制造)yuánshǐ shèjì zhìzào | A factory that has already designed a product and will put your name on it. | Original design manufacturer: the factory owns the design and often the tooling, sells it to several brands and usually sets a minimum order. Fast and cheap to start; weak on differentiation and control. |
| OEM (contract build) | OEM代工OEM dàigōng | You bring the design, the factory builds exactly what you drew. | In a supply contract, the factory builds to the customer's design, and the customer owns the design and usually the tools it paid for. Distinct from "robot OEM", the maker that sells robots under its own brand (本体厂商). |
| White label | 贴牌tiēpái | A finished product made by someone else, sold with your brand on it. | Rebadging a maker's standard product under the seller's brand, with little or no design change. Common in service robots and in after-sales work done under a distributor's name. |
| Sample | 样品 / 打样yàngpǐn / dǎyàng | One or a few parts a supplier makes so you can check them before ordering more. | Pre-production parts from a supplier for fit, function and quality checks, often charged. 打样 (dǎyàng) is the verb, to make a sample. |
| Prototype | 样机yàngjī | An early working robot built to test the idea, usually by hand and at high cost. | A functional pre-production unit, typically machined and printed, used to prove the design before tooling. Prototype-to-production is where most hardware schedules slip. |
| Mass production | 量产liàngchǎn | When the moulds are made and the factory builds the same robot over and over at a low price. | Production on hard tooling at stable yield and volume, with the lowest unit price. Chinese makers announce 量产 early; ask for delivered units, not the announcement. |
| Small batch | 小批量xiǎo pīliàng | Building tens or hundreds of units to prove the factory can make them before going big. | Pilot production between prototype and mass production, often on soft tooling or machined parts, priced higher per unit. |
| MOQ | 起订量qǐdìngliàng | The smallest order a supplier will accept. | Minimum order quantity: the smallest run a supplier will quote, set by setup cost and material lots. Negotiable against a credible forecast or a higher unit price. |
| Certification | 认证rènzhèng | A test lab checks your robot against the rules and gives it a stamp so you are allowed to sell it. | Type testing of a design against standards before sale: CE, FCC, UL, RoHS, UN38.3 for batteries, national radio approvals (IMDA Singapore, SIRIM Malaysia, SDPPI Indonesia) and robot standards such as ISO 13482 and SS 713. TÜV SÜD PSB and SGS test robots in the region. |
| NRE | NRE费用(一次性工程费)yīcìxìng gōngchéng fèi | The one-time engineering bill a supplier charges to set up your job, before any parts are made. | Non-recurring engineering: one-off design, setup, fixture and qualification charges, separate from tooling and the per-unit price. Negotiate whether it is refunded against volume. |
| DFM | Designing a part so the factory can make it cheaply and easily, before you order the moulds. | Design for manufacture (with design for assembly, DFA): rules and scoring, such as Boothroyd and Dewhurst's three questions per part (does it move, must it differ in material, must it come apart), that cut part count, tooling and assembly steps. | |
| Learning curve | Every time you make twice as many, each one gets cheaper by about the same percentage. | Unit cost falls by a fixed fraction with each doubling of cumulative volume (Wright's law). Leju's actuated axis fell from US$433 to US$202 over 100x volume: about 11% per doubling, an 89% curve (estimate from corematter data). |
Sources for this chapter
- chinabizinsider, Unitree G1 teardown (China Post Securities), 29 Apr 2026: chinabizinsider.com
- BigGo Finance, SemiAnalysis Unitree cost estimate, 9 Jun 2026: finance.biggo.com
- 21jingji, Unitree prospectus and IPO, 19 Aug 2026: 21jingji.com; 36Kr prospectus deep dive: 36kr.com
- Corematter, the US$202 humanoid joint, 11 Sep 2026: read.corematter.com
- CryptoBriefing, Optimus build cost, 6 Jun 2026: cryptobriefing.com; TechDigest, excluding Chinese parts, Feb 2026: techdigest.tv
- 36Kr, humanoid value chain and roller screw capacity: eu.36kr.com
- 163.com, harmonic reducer market shares: 163.com; OFweek, CCID 6-axis report: mp.ofweek.com
- eastmoney, Leadshine capacity: caifuhao.eastmoney.com
- CNX Software, NVIDIA Jetson price increase, 22 Jul 2026: cnx-software.com
- Figure, ramping Figure 03 production, 29 Apr 2026: figure.ai; Figure 03 battery, 17 Jul 2025: figure.ai
- Agility, SEC Form 425 (RoboFab), 26 Aug 2026: sec.gov
- Electrek, Xpeng IRON production line, 7 Sep 2026: electrek.co
- Singapore National Robotics Programme, standards: nrp.gov.sg; TÜV SÜD robotic safety: tuvsud.com; SGS robotic products testing: sgs.com
- Unitree L2 lidar: unitree.com
- Materials primer (research corpus, Jul 2026): material properties, process costs, crossover model, tolerance costs, margin map, failure modes. Industry report (research corpus, 28 Jun 2026): BOM shares, supply-chain layers.