Why the Harmonic Reducer Is the Soul of a Humanoid Joint

By Theodore Li, Technical Director, HONPINE. Adapted from an explainer by the Chinese reducer engineer known as “Old Ou” (做减速机的老欧), who has spent more than a decade in precision gearbox manufacturing.

Fourteen. That is the number worth starting with. Of the 28 rotary actuators in Tesla’s Optimus, 14 use the same recipe: frameless torque motor, harmonic reducer, torque sensor, encoder. They sit in the shoulder, the wrist, the hip and the torso.

Why that combination, in that place, and not something cheaper? The answer is a mechanical argument, not a purchasing one — and it comes with a bill that gets paid by one very thin-walled part.

The mechanism: an ellipse instead of a gear pair

Start with the contradiction every joint has to solve. A robot joint motor spins at several thousand rpm but delivers a fraction of a newton-metre. To lift a few kilograms, you must trade speed for torque, at a large ratio, inside a space about the size of a fist.

Conventional gearboxes do this by putting hardened teeth against hardened teeth. The harmonic reducer does it by bending metal.

It has three parts and only three:

• a wave generator — an elliptical cam with a thin ball bearing race on its outside
• a flexspline — a thin-walled, cup-shaped elastic gear with external teeth
• a circular spline — a rigid ring gear with internal teeth

The wave generator is pressed into the flexspline and forces it into an ellipse. At the two ends of the major axis the flexspline teeth are fully meshed with the circular spline; at the minor axis they are completely disengaged. The flexspline has exactly two fewer teeth than the circular spline, so every full turn of the wave generator walks the flexspline backward by two teeth.

The clever part is not the ratio. It is how many teeth carry the load at once. A double-wave harmonic unit engages more than 30% of its total tooth count simultaneously, against one or two pairs — roughly 2% to 7% — for a spur gear pair. Load spread over dozens of teeth means each tooth sees less force, the drive is smoother, and the positioning is more repeatable: ±10 arcsec repeatability and under 1 arcmin of backlash are routine claims, in a single stage, at ratios from 50:1 to 250:1.

Why humanoids cannot route around it

An industrial robot lives behind a fence and can weigh two tonnes. A humanoid has to work in the same corridor as a person, which means its joints cannot be bulky, cannot protrude, cannot whine, and cannot weigh several kilograms each.

That single constraint eliminates most of the gearbox catalogue. The harmonic reducer keeps its input and output on the same axis, so the envelope is a cylinder rather than an offset housing. Single-stage ratio reaches 250:1 without a second stage, which removes both a gear set and the backlash that comes with it.

At CIIF 2025 in Shanghai, one supplier showed a double-circular-spline harmonic unit claimed to weigh 0.32 kg — 68% lighter than a conventional harmonic of the same rating. That is a company figure and has not been independently verified at volume, so treat it as a direction of travel rather than a specification.

The three reducer families are not grades of the same product. They occupy different joints:

Pair a harmonic reducer with a frameless torque motor and you get the joint module that most humanoid programmes now buy as a unit. Current mainstream modules advertise peak torque density around 100 Nm/kg and power density up to 300 W/kg.

The other side of the bill: the flexspline

Here is the part that sales decks leave out. The flexspline is deformed into an ellipse twice per revolution. Every point of that thin wall is loaded in alternating tension and compression, millions of times, for the life of the joint.

Fatigue fracture of the flexspline accounts for more than 60% of all harmonic reducer failures. It is not an edge case; it is the dominant failure mode by design.

Domestic units typically run 6,000 to 8,000 hours. Japanese product from Harmonic Drive Systems reaches above 10,000 hours. Even where lifetimes overlap, the fatigue performance of domestic flexsplines still trails Harmonic Drive by 20% to 30%.

Part of that gap is metallurgy. Harmonic Drive uses a V-Ti-Nb micro-alloying design; most domestic flexsplines use V-Ti only. Niobium refines the grain structure, and a finer grain resists crack initiation under repeated deformation. Measured against the Japanese material, domestic flexspline structures are coarser by 1.0 to 4.0 grain-size grades, and less consistent — a grade spread of up to 3.0.

There is one frequently cited failure case that deserves a caveat. A 32-120 harmonic unit cracked its flexspline after 23 days running at 3,000 rpm — 150% of rated speed — under a 15 kg load, 110% of rating. The failure analysis by Yuan Feng et al., published in Metal Working, records the case and classifies it as low-cycle fatigue, noting the unit’s normal service interval should have been 60 days or more. That is a deliberately abused unit in an extreme corner of the envelope, not a typical lifetime. It is useful as evidence of how the flexspline fails, not as evidence of how long it lasts.

Vacuum makes all of this worse. Without atmospheric pressure to help lubricant reach the gap between flexspline and wave generator, wear accelerates sharply. A humanoid joint may flex tens of thousands of times a day, and one failed joint stops the whole machine — the single risk OEMs are least willing to accept.

The substitution story, with the caveats attached

Harmonic Drive Systems has held more than 40% of global harmonic capacity for years. That is changing at the margin.

Suzhou Leaderdrive (绿的谐波) developed its own P-tooth profile: precision within 10 arcsec, service life past 20,000 hours, at a lower price than the equivalent Harmonic Drive unit. In 2025 it held roughly 20% of the global market — second globally, first in China — and is in the Optimus supply chain.

Demand projections are aggressive. Laifu Harmonic estimates China’s robotics harmonic reducer market at about ¥1.3 billion in 2025, rising to ¥13.4 billion by 2030, a compound annual growth rate near 59.8%.

But a capacity gap is not automatically a domestic opening. Harmonic Drive’s capacity was 3.06 million units a year in 2024, projected to reach 3.47 million in 2026 — an expansion of only about 13%. That looks conservative. Capacity utilisation, however, is the variable that decides whether conservative expansion means shortage. If the installed base is not running flat out, a 13% expansion is not a bottleneck.

The window for domestic suppliers is real. Whether it gets used depends on three things that have nothing to do with marketing: progress in flexspline metallurgy, consistency at volume, and credible life validation.

FAQ

Why does a humanoid joint need a reducer at all?
Because joint motors spin at thousands of rpm with very little torque, while the joint needs low speed and high torque. The reducer performs that conversion, and it also sets backlash, heat and service life — which is why changing the reducer changes the control architecture.

What ratio does a harmonic reducer achieve in one stage?
Typically 50:1 to 250:1, with input and output on the same axis, which is what makes the small envelope possible.

Is a harmonic reducer more precise than an RV reducer?
On positioning and backlash, yes: harmonic units reach ±10 arcsec with under 1 arcmin of backlash, and cycloidal and RV designs are measured below that. RV wins on load capacity and stiffness, which is why it still dominates industrial robot base axes.

Why is backlash so important in a humanoid?
Every arcminute of backlash is a dead zone the controller must model and compensate. In a wrist holding a tool, backlash becomes visible positioning error; in a leg, it becomes a limit cycle the whole-body controller has to damp.

Which part fails first?
The flexspline, in more than 60% of cases. It is elastically deformed twice per revolution, so the failure is fatigue at the tooth root — a design characteristic, not a manufacturing defect.

Can domestic harmonic reducers replace Harmonic Drive today?
For most humanoid joints, yes on precision and cost; the open questions are fatigue life in the 6,000–8,000 hour band versus 10,000-plus, and batch-to-batch consistency of flexspline material.

Source: adapted with permission of the argument from the WeChat technical account 做减速机的老欧 (Zuo Jian Su Ji De Lao Ou), “Why the harmonic reducer is the soul of a humanoid joint”. Figures on Optimus actuator count, flexspline fatigue share, material composition, pricing and capacity are as reported in that article; the CIIF 2025 weight claim is the supplier’s own and remains unverified at volume. Published by Latest Robot News.