Gears and screws

A robot motor spins fast and pushes weakly. A robot joint has to move slowly and push hard. The reducer between them trades one for the other, and in a humanoid that trade is worth more money than the motor itself: about 36% of a rotary joint's cost is the gearbox and about 64% of a linear joint's cost is the screw estimate. Learn the five mechanisms on this page and you can read any joint spec sheet in Shenzhen and know what the maker chose to be good at.

Why a joint needs a gearbox

An electric motor makes power, and power is torque times speed. A small motor of the kind that fits inside an elbow makes its power as speed: thousands of revolutions per minute at a torque you could stop with your fingers. An elbow lifting a 5 kg box needs the opposite, a few revolutions per minute at a torque that would snap your wrist. A gear pair converts one into the other. If a small gear with 12 teeth drives a big gear with 48 teeth, the big gear turns once for every four turns of the small one, and it pushes four times as hard, minus friction.

That multiplier is the gear ratio. Here is the arithmetic. Say a motor spins at 3,000 rpm and makes 0.5 N·m. Put a 50:1 reducer on it and the output spins at 60 rpm with up to 25 N·m, less whatever the gears lose as heat. At 80% efficiency you get 20 N·m. Stack two stages and the ratios multiply: 4:1 followed by 4:1 is 16:1.

A single spur pair would need a joint gear 160 times the diameter of the motor gear to reach 160:1. So robot joints use packed mechanisms that reach high ratios in a small, round, hollow space. More than 90% of rotary joint modules use one of two, harmonic or planetary estimate, with cycloidal reducers for heavy, shock-loaded joints and screws wherever a joint pushes in a straight line.

Harmonic drive: two teeth of difference

The harmonic drive, also called a strain-wave gear, gets a huge ratio from three parts and one trick. The circular spline is a rigid steel ring with teeth on the inside, bolted to the robot's frame. Inside it sits the flexspline, a thin steel cup with teeth on the outside and two fewer teeth than the ring. Inside the cup the motor spins the wave generator, an oval cam on a thin flexible bearing. The oval squeezes the cup into an ellipse, so the cup's teeth mesh with the ring only at the two ends of the long axis.

Turn the oval once and the mesh points travel all the way round the ring. Because the cup has two fewer teeth, after one lap it has fallen two teeth behind: it has turned backwards by 2 ÷ (cup teeth) of a revolution. A cup with 100 teeth in a 102-tooth ring gives 50:1. A cup with 160 teeth gives 80:1, and a 320-tooth cup gives 160:1, all in one flat stage. Because teeth are always engaged on both sides with preload from the oval, there is almost no play.

That's why harmonic drives sit in arms, wrists and shoulders, where you want precise positioning in a small, light package. Optimus Gen 1 used 14 rotary actuators with harmonic-type reducers verified, and the typical humanoid rotary joint is a frameless motor, a harmonic reducer, a torque sensor and an encoder stacked around a hollow shaft (see Motors and joints). The value in the part sits in the hard pieces: the flexspline, the wave generator and final assembly and test are more than 70% of a harmonic reducer's unit cost estimate.

Planetary gearsets and quasi-direct drive

A planetary gearbox packs gears like a solar system. The motor turns the sun gear in the middle. Three or four planet gears mesh around it and also mesh with a ring gear that has teeth on the inside. Hold the ring still and the planets roll around the inside of it, carrying the planet carrier with them. The carrier is the output. With the ring fixed, the ratio is 1 + ring teeth ÷ sun teeth.

One planetary stage gives 3:1 to 10:1 estimate. That low ratio is a feature. The joint stays easy to push from the outside, a property called backdrivability, so the motor current alone tells the controller how hard the leg is hitting the ground. That's the logic of quasi-direct drive: a big, flat, high-torque motor with a 6:1 to 10:1 planetary, used by Unitree and the MIT Cheetah lineage. You give up some torque density and gain a leg that absorbs impacts and senses force without a separate torque sensor.

Unitree's G1 shows what this costs at Chinese scale. Its small joint is a 4-in-1 unit (motor, planetary reducer, encoder, driver) around a hollow shaft; the planetary stage costs about CNY 300, the motor CNY 200 and the encoder CNY 200, and the whole small joint about CNY 1,000 (US$140) estimate. Unitree designs its own planetary tooth profile to survive leg shock verified, per its IPO prospectus.

Cycloidal and RV reducers: many teeth share the hit

A cycloidal drive replaces teeth with lobes. The input shaft carries an off-centre cam. The cam makes a wavy disc wobble in a small circle inside a ring of fixed pins. The disc has one fewer lobe than there are pins, so each full wobble walks it back by one lobe. Pins sticking through oversized holes in the disc pick up that slow rotation and ignore the wobble. With 10 lobes and 11 pins the ratio is 10:1, and it runs the opposite way to the input.

An RV reducer is a cycloidal drive with a planetary stage in front: the motor turns a small planetary set, which spins two or three crank shafts, which wobble two cycloid discs set half a wobble apart so they balance each other. It is heavy and expensive, and it is the standard at the base, shoulder and elbow of industrial arms, where one joint carries the whole arm plus payload. Nabtesco of Tokyo holds about 60% of RV reducers for medium and large industrial robots estimate, and Chinese RV makers supply more than 30% of the robots assembled in China estimate.

In humanoids, cycloidal units show up where the load is repetitive and heavy. Agility's Digit 5, unveiled 15 Sep 2026, reportedly revised its legs around cycloidal actuators for repeated lifting unverified. The industry report's summary rule holds as a first pass: harmonic for precise arms and hands, cycloidal for shock-loaded hips and knees, screws for straight-line thrust.

Screws: turning spin into a push

Some joints are easier to drive by pushing than by turning. A human knee works that way: a muscle pulls in a straight line across the joint. A linear actuator copies it. A motor turns a threaded shaft, a nut rides along the thread, and the nut pushes a rod that pushes the limb. One turn of the shaft moves the nut one lead, for example 5 mm. That is a reducer too: a small motor torque becomes a large straight-line force.

A ball screw puts steel balls between the shaft and the nut so they roll instead of sliding. A planetary roller screw replaces the balls with small threaded rollers that orbit the shaft like planets. Each roller touches both the shaft and the nut along many thread lines, where a ball touches at one point, so a roller screw carries 3 to 10 times the load of a ball screw the same size and survives shock estimate. That is why Tesla put 14 of them in Optimus verified, Boston Dynamics' electric Atlas reportedly uses them unverified, and Kepler's K2 mixes roller-screw linear actuators with rotary joints maker release.

Why the roller screw is the bottleneck

A roller screw is a few pieces of hardened steel with threads ground to a few microns. Three things make it slow to scale. The alloy: about 70% use titanium or chrome-moly steels, only about five foundries make aerospace-grade stock, and steel lead times run 6 to 9 months unverified. The machines: threads are finished on precision thread grinding machines, mostly imported, and grinding capacity is the rate limit. The know-how: heat treatment and tolerancing that incumbents built over decades. In 2023 roller screws were only 4% of China's CNY 25.7 billion screw market, and foreign makers held 85% of China's roller screw sales estimate.

The price depends on who is quoting it, and the gap between sources is large enough that it matters exactly what is being priced.

The likely reconciliation: the US dollar figures describe low-volume aerospace screws or whole actuators, and CNY 1,500 is a volume price for a screw alone estimate. Chinese capacity is now arriving in bulk. Wuzhou Xinchun is building for 980,000 roller screw sets plus 2.1 million micro ball screws; Beite plans 2.6 million sets a year in Kunshan plus 800,000 in Thailand; Hengli expects its screw line to exceed CNY 3 billion of output estimate. Against that, UBS projects about 30,000 humanoids shipped worldwide in 2026 estimate. Capacity running ahead of demand points to a price war in screws estimate, the same pattern harmonic reducers already show.

Belts, cables, tendons and linkages

Every kilogram at the end of a limb costs you twice: the motor has to lift it, and it swings like a hammer when the robot falls. So designers move motors up the limb toward the body and send the motion down with something light.

• Timing belts are toothed rubber belts with steel or fibre cords. They carry rotation from a motor to a joint a short distance away, cheaply and quietly, with a small extra ratio from the pulley sizes. They stretch a little, so they add compliance and some position error.
• Tendon drives pull joints with cables, like the tendons in your forearm that move your fingers. 1X builds NEO around tendon drive, which keeps its joints light, quiet and easy to push by hand; NEO weighs 30 kg verified. Optimus Gen 3's hand reportedly uses about 25 forearm actuators pulling tendons unverified. Cables stretch, fray and need tensioning, so they trade maintenance for lightness (see Hands and touch).
• Linkages are rigid rods and levers. A motor mounted high on the thigh can drive the knee or ankle through a push rod, and two rods from calf-mounted motors can tilt an ankle in two directions. Agility's Digit inherits its backward-bending bird legs from the Cassie research robot verified, which came out of walking-physics research. Linkages are stiff and efficient, but the ratio changes as the joint moves, which the controller has to model.

Seven properties that decide the choice

No reducer wins on every axis. A joint sold as "better" is better on some of these seven and worse on others.

Reflected inertia deserves a picture, because it is the reason legged robots moved away from high ratios. A motor rotor spinning N times faster than the joint stores N² times more kinetic energy per unit of joint speed. When the foot hits the floor, the gear teeth have to stop that energy.

Which gear goes where

Ratios for harmonic, planetary and screws come from the sources for this chapter; the other cells are relative rankings. Here is what the named robots chose, as far as it's public.

Who makes them and what they cost

For decades two Tokyo firms owned precision reducers: Harmonic Drive Systems (HDSI) in strain-wave gears and Nabtesco in RV. Together they still hold about 55 to 60% of global harmonic revenue estimate, and HDSI alone about 60% of the global harmonic market by one count estimate. Inside China the picture has flipped, and this is the clearest case of domestic substitution in the robot supply chain.

Two things are happening at once. Chinese harmonic reducers are 30 to 50% or 40 to 60% cheaper than Japanese ones, depending on the source contested, while HDSI's custom or micro units sell for CNY 7,000 to 8,000 estimate. And margins are falling as capacity floods in: Leaderdrive shipped about 500,000 units in 2025 with capacity approaching 1 million in 2026 estimate, its H1 2026 revenue grew 38.6% to CNY 349 million, but its operating cash flow was negative CNY 17 million verified. A reported "3.5 million units of locked orders for 2026" doesn't fit a 1 million unit capacity and is probably a translation or unit error contested.

One structural point matters for geopolitics. Magnets live in the motor, so reducers and screws are less exposed to rare-earth export controls than motors are. Their exposure is to precision steel, heat treatment and grinding machines. Integrated joint modules bundle both and inherit the magnet risk.

How transmissions fail

Gears don't fail at random. They wear in a few predictable ways, and the signs show up in data the robot already collects.

• Flexspline fatigue. Every input turn bends every point of a harmonic drive's thin cup twice. A 100:1 joint doing one output turn flexes the cup 200 times. Over millions of cycles a crack starts at the tooth root, usually after shock loads or overheating. The joint then loses torque or skips a tooth.
• Grease breakdown. Grease thins with heat, gets pushed out of the mesh and picks up wear particles. Fanuc's Zero Down Time service tracks grease change timing across more than 40,000 connected arms verified. Humanoid joints run hotter and are packed tighter, so expect shorter grease life.
• Backlash growth. Worn teeth and bearings add play. You see it as a joint that overshoots, hunts around its target or clunks when it reverses. It creeps in slowly, so a fleet operator who logs position error per joint can see it long before a customer does.
• Broken small gears from control software. A lab running reinforcement learning policies on a Unitree G1 reported a broken left wrist pitch gear after violent motions in unfamiliar states verified owner report. In humanoids, the controller causes a lot of the gear damage.
• Roller screw wear and contamination. Dirt in the nut and lost lubrication pit the threads. Linear actuators sit in the legs, close to floor dust and water.

Terms in this chapter