BLOGS

In brief: A humanoid robot is not one machine that needs one reducer. It is dozens of joints with different torque, precision, shock and cost personalities — and 2026 supply-chain data shows the three main reducer families splitting even further apart: harmonic drives running away with the upper body, planetary gears taking the volume positions in the legs, and RV (cycloidal) units holding the heavy-load niches with the deepest localization gap. Here is how the joint map sorts the three technologies — and what that means when you specify an integrated joint module.

The Joint Map Decides — Not the Datasheet

Industrial robots settled this question decades ago with a simple rule of thumb: RV reducers for joints 1–3 (base, shoulder, upper arm) where torque and rigidity dominate, harmonic reducers for joints 4–6 (wrist) where precision and compactness dominate. Humanoid robots inherited the logic, then scrambled it. A Tesla Optimus-class machine carries roughly 28 body joints and more than 50 degrees of freedom once the hands are counted — a population so heterogeneous that no single transmission family can serve it efficiently.

The 2026 shipping data makes the split measurable. Interim disclosures from one of China's two leading harmonic suppliers show that sub-52 mm miniaturized harmonic units — the class used in wrists, spines and dexterous hands — grew from 5.1% to 10.6% of shipments in a single year, with first-half harmonic shipments up 112% year over year. Industry estimates put the harmonic count per humanoid at 10–20 units today, rising as dexterous-hand designs add micro-actuated fingers. The upper body is becoming harmonic territory at an accelerating pace.

Harmonic Drives Own the Upper Body

The harmonic (strain wave) reducer wins its territory on geometry. Its thin-walled flexspline achieves a high single-stage reduction ratio inside a very short axial envelope, which is exactly what a shoulder, elbow, wrist or waist joint needs: flat, light, easy to route cables through, and stiff enough for positioning tasks. Backlash is near zero, and humanoid-specific products now quote repeatability in the ±10–15 arcsec class.

The physics that disqualifies it elsewhere is fatigue. The flexspline transmits torque through elastic deformation of a thin steel cup, so sustained shock loads and abuse cycles are its worst enemy. Industry selection guidance for 2026 draws the line explicitly: harmonic stages should stay within roughly 3× rated peak torque for no more than about 100 ms, and most vendors discourage them for the main hip and knee load paths of adult-scale humanoids. The upper body, where loads are lighter and motion is more about precision than impact, is where the technology is unmatched.

Planetary Gears Take the Volume Positions

Planetary reducers are the oldest, most mature transmission in the stack — and in humanoid robots they are winning by cost-per-newton. Industry surveys estimate planetary-based integrated joint modules account for the large majority — on the order of 80% — of humanoid joint-module units, driven by cost-optimized lower-body designs: several volume-oriented Chinese humanoid makers build their hips and knees on planetary modules precisely because the manufacturing process is mature, yield is high, and delivery consistency suits mass production.

Two engineering reasons complete the case. First, efficiency: a planetary stage wastes less torque as heat, which matters in legs that work all day on battery power. Second, shock tolerance: spur-gear meshes tolerate landing impacts better than a strained flexspline. The trade-off is backlash — planetary stages accumulate more of it, which is why precision-critical axes avoid them. The fast-growing quasi-direct-drive (QDD) architecture — a low reduction ratio paired with a high-torque-density motor — builds on the planetary baseline and, per industry monitoring, has reached roughly one-third penetration in mid- and light-load segments, prized for force-control responsiveness in ankles and wrists.

RV Reducers: the Rigidity Reserve — and the 25% Problem

The RV (rotary vector, cycloidal) reducer is the rigidity reserve of the joint world. Its two-stage cycloidal design spreads load across multiple tooth contacts, delivering the highest torsional stiffness in the group, shock overload capacity around 5× rated torque, and L10 design life at or above 10,000 hours. For heavy base joints, waists and industrial-grade hips, nothing else holds position under sustained impact as well.

Here is the paradox the 2026 data exposes: in harmonic reducers, domestic supply has taken an estimated 60–90% of the humanoid-specific segment, with domestic products carrying 40–60% of the cost of imported equivalents. In RV reducers, domestic share remains around 25%, with the Japanese incumbent still holding a commanding lead. The bottleneck is not geometry — domestic cycloidal wheels are already machined to ±0.005 mm class precision — but batch-to-batch consistency and lifetime validation under real duty cycles, which only volume production and years of field data can prove. We covered what a cycloidal-first leg design looks like from the customer side in our analysis of Agility's Digit 5 leg redesign.

The 2026 Snapshot: One Table, Three Races

The three families, their territories and the supply-side race, in one view (shares are industry estimates and vary by source):

One more integration trend is worth tracking: industry monitoring suggests more than 87% of humanoid integrated joint modules now ship with an on-board torque sensor and dual encoders as standard — the reducer is becoming one component inside a sensing-rich actuator unit, not a separately specified part.

What to Lock Down Before You Choose

1. Rate the joint by rated torque, not peak. Peak numbers are transient-event ratings; sustained duty at rated torque is what decides thermal survival.
2. Budget for shock explicitly. Match the transmission to the impact profile: roughly 3×/100 ms tolerance on the harmonic side versus 5×/200 ms on the RV side is the current industry yardstick.
3. Set backlash and stiffness per joint, not per robot. A wrist wants arc-seconds; a hip wants newton-meters per radian of stiffness; the same spec sheet cannot serve both.
4. Hold an efficiency floor. Around 75% for harmonic stages and 80% for RV stages keeps heat and battery budgets honest in leg-class duty cycles.
5. Demand fatigue data for L10 life. A 10,000-hour claim without load-spectrum-backed testing is a slogan, not a specification.
6. Specify integration, not parts. Motor, reducer, driver and dual encoders engineered as one module — with hollow routing and a standard industrial bus — removes the integration errors that spec-sheet math never shows.

The Ti5ROBOT View

The joint map is the specification. Our Lightweight Joint Module series is built for the harmonic-class precision positions — compact envelopes, dual absolute encoders, hollow routing — while our humanoid robot actuator line covers the shock-tolerant and higher-torque positions where different transmission physics earn their place.

That is also the theme of our earlier comparison of harmonic and planetary routes for joint modules, and of our breakdown of the cycloidal bet in Agility's Digit 5 leg redesign: the physics of the joint decides, integration quality decides, and no single reducer pretends to fit everywhere.

The Bottom Line

The 2026 divergence is not a battle with one winner — it is a sorting. The joints sort the reducers: harmonic up top, planetary underneath, RV where rigidity is non-negotiable. The supply chain sorts the suppliers: one family already localized past 70%, another still stuck at 25% for reasons of consistency rather than capability. The teams that win are the ones that specify transmission per joint, verify fatigue and shock data per vendor, and buy integration quality instead of gear ratios.

Source: Ti5ROBOT Blog, September 2026. Based on listed-company interim disclosures (2026 first-half shipment, utilization and backlog data from leading Chinese harmonic suppliers) and industry monitoring of humanoid joint-module localization and published selection thresholds. Figures marked as estimates vary by source; single-source figures are attributed as such. No pricing information is included by design.