BLOGS

In early October 2026, the long-simmering question — should a humanoid robot's joints use a reducer at all? — spilled out of engineering forums and into the public markets. A direct-drive module maker listed in Hong Kong, a rumored architecture change circulated around Tesla's next humanoid, and suddenly "direct drive vs. geared joints" became a boardroom topic. Behind the noise sits a genuinely important engineering decision that every humanoid program will face: which joints benefit from removing the gearbox, and which ones will keep one for a decade or more. This article maps the physics, the economics, and the likely endgame: not a winner-take-all victory, but a per-joint hybrid architecture.

The Debate, Stripped of the Hype

First, what actually happened. On September 29, 2026, Benmo Technology (06731.HK), a direct-drive power module maker, listed on the Hong Kong Stock Exchange main board and was widely described in Chinese financial media as the world's first listed "pure direct-drive" module company. The same day, it signed a strategic cooperation agreement with China Rare Earth Group's innovation subsidiary (China Rare Earth Group Innovation Technology Co., "Zhongxi Kechuang") covering rare-earth magnet materials for robot power modules — per Securities Daily and the company's own announcements.

Separately, market rumors claimed Tesla's next-generation Optimus would shift major joints to direct drive. No official source has confirmed any such change; a Huatai Securities research note reportedly framed the lower-body reducer and lead-screw scheme as the segments with the highest probability of revision. That is speculation, not fact — but the rumor's persistence is itself informative: the market has already assumed the joint actuator architecture is on the table.

Source: OFweek Robotics, "Optimus V4 rumors stir Hong Kong stocks: the route showdown between robot direct drive and reducers," October 6, 2026; Securities Daily, September 30, 2026.

Why Direct Drive Kept Losing on Torque

The core problem has never been controversial. A gearbox is a torque multiplier: a harmonic reducer with a 50:1 or 100:1 ratio turns a small, fast, efficient motor into a compact joint with 50–100 times the output torque. Remove it, and the motor itself must supply all the torque. For the same output torque, that means a dramatically larger stator diameter, stronger magnets, and more copper — mass that lands exactly where humanoids can least afford it, at the distal end of the limb.

Three second-order problems follow:

• Heat. Without a reducer, the motor must generate full joint torque at low speed, where cooling is hardest and current — the main source of resistive heating — is highest.
• Efficiency. Low-speed, high-torque operation sits in a poor region of most motor efficiency maps.
• Precision. Encoders must resolve angle directly at the output shaft with no gear reduction helping the resolution on the motor side; position accuracy at the joint depends entirely on the feedback stack.

This is why direct drive historically survived only in niches — haptic interfaces, some legged research platforms, and high-speed light-load pickers — while harmonic and cycloidal gears owned anything that needed real torque in a small envelope.

What Mass Production Actually Changed

The interesting story of 2026 is not that direct drive got better on the slide deck. It is that someone ran the manufacturing learning curve at consumer-robot scale. According to Frost & Sullivan data cited in Benmo Technology's listing materials and reported by multiple outlets, the company's direct-drive power module shipments grew from roughly 186,000 units in 2023 to about 2.14 million in 2024 and over 8.5 million in 2025 — the first and, so far, only module maker to cross five million units in a year. In the first half of 2026 it shipped roughly 5.7 million more, largely into consumer robots: floor cleaners, lawn mowers, and similar high-volume, moderate-load products.

That trajectory matters for a simple reason. Consumer robots are an unfashionable but brutal school: millions of units force every cost driver — magnet grade, motor topology, lamination stacking, automated winding, tolerance control — into a disciplined optimization loop. The company disclosed that a 2023 redesign of magnet material and motor topology raised torque density by more than 1.5× compared with its earlier permanent-magnet designs (a company-sourced figure). Revenue grew from about 17.5 million RMB in 2023 to roughly 282 million RMB in 2025 along the way.

The rare-earth cooperation adds a second lever. Direct-drive machines are, more than anything else, magnet machines: low speed and high torque in a small volume are bought with high-energy NdFeB and clever magnetic circuit design (concentrated windings, Halbach-like arrays, flux-concentrating rotor structures). Binding a magnet supplier with upstream rare-earth resources into the product development loop shortens the material-to-module path — and, as several Chinese analysts noted, shifts the comparative advantage of the direct-drive route toward whoever controls magnet production at scale.

The Physics Gears Still Win

None of this means the gearbox dies. Four properties keep geared joints — and increasingly, specific gear types per joint — in the picture:

• Impact tolerance. Legs take impacts. A reducer's gear train presents a lower effective inertia to the output and a mechanical impedance that protects the motor rotor; a direct-drive rotor coupled 1:1 to a stomping foot absorbs every shock electrically or mechanically, immediately.
• Stiffness and reflected inertia. For position-holding under load, the gear ratio raises stiffness and reduces the load inertia reflected to the motor — which is why heavy-duty industrial joints still favor cycloidal (RV) stages.
• Thermal margin. Geared joints let the motor run at higher speed and lower current for the same torque, moving the operating point to a cooler, more efficient region.
• Envelope. A 100:1 harmonic stage still delivers more torque per millimeter of actuator diameter than any motor alone. For wrist and shoulder joints in a human-scale frame, that envelope advantage is decisive.

The honest reading of the table: no single column dominates. Every row is a different winner — which is precisely why the "one drive technology to rule them all" framing keeps failing.

The Endgame Looks Hybrid, Not Winner-Take-All

Two data points from 2026 make the convergence argument better than any essay could. First, industry white papers and broker analyses have converged on a per-joint division of labor: harmonic stages for upper-body precision joints, quasi-direct-drive or planetary schemes for the lower body's high-dynamic joints, with cycloidal stages where shock and payload dominate. Second — and more telling — the world's most prominent direct-drive module maker itself unveiled a heavy-load geared joint module family at the 2026 World Robot Conference, explicitly targeting shock-heavy scenarios that pure direct drive cannot yet cover. When the direct-drive champion starts selling reducers, "direct drive vs. geared" has clearly become "which reducer where, and which joints can drop the gear entirely."

The economic subtext also reshapes valuations, not just designs. Analysts estimate the harmonic reducer accounts for roughly 36% of a rotary actuator's cost (a figure popularized by Bernstein's analysis of humanoid actuator BOMs and widely cited in Chinese coverage). If high-load positions migrate toward cycloidal stages or lower-ratio planetary schemes, and light positions go gearless, that share compresses — the "units × per-unit reducer value" growth story that reducer makers have been trading on becomes non-linear.

What This Means for Actuator Buyers

For engineering teams specifying humanoid or manipulation platforms today, the route debate resolves into a per-joint checklist rather than a slogan:

• Map torque, speed and duty per joint first. A wrist roll joint at 20 Nm continuous and a hip joint at 200 Nm peak with impact live in different technology quadrants; forcing one answer across the body wastes either performance or mass.
• Ask where the encoders sit. Gearless joints shift the precision burden entirely onto the feedback stack — dual-encoder architectures (motor-side plus output-side) are the minimum for closed-loop accuracy claims. We covered the accuracy ceiling set by encoder architecture in Inside the Integrated Joint.
• Demand shock ratings, not just torque curves. For lower-limb joints, ask how the module handles impact loads and what its overload envelope looks like for seconds, not milliseconds.
• Check backdrivability requirements honestly. Force-controlled and compliant manipulation want low reflected inertia; a gear train in the wrong place turns a force-control problem into a control-fight problem. The sensor side of this trade-off is dissected in The Force-Sensing Stack.
• Weight the supplier's production maturity, not its demo videos. The 2026 lesson is that a module designed for millions of consumer units arrives with cost and reliability discipline that a lab-built equal-spec module cannot match.

This is also the design philosophy behind Ti5ROBOT's joint products. Our Lightweight Joint Module family packages motor, reducer, dual feedback and drive electronics into integrated rotary joints sized for humanoid and light industrial arms — for example the 30–40 class joint module — while our Humanoid Robot Actuator line targets the higher-torque positions where the gear stage remains the right answer. The point is not to pick a camp; it is to pick the right transmission per joint and own the integration.

Signals Worth Watching

The debate will be settled by evidence, and three observable signals matter more than any roadmap:

• Design wins in shipped humanoids. Which transmission scheme actually lands in the joint BOM of the next volume humanoid generation — not the prototype, the production run.
• Magnet-to-module industrialization. Whether the rare-earth cooperation produces qualified magnet materials in production modules, closing the material cost loop for direct drive.
• Revenue mix beyond consumer robots. Whether direct-drive module makers convert consumer-robot cost discipline into humanoid design wins, visible in their non-consumer revenue share over the next two fiscal years.

One caveat for readers tracking this through capital markets: the listed direct-drive maker's post-IPO run-up reflects scarcity and narrative as much as fundamentals, and third-party valuation screens already flag its profitability as unproven. The technology curve and the stock curve are different objects.

The gearbox question, in the end, is a cost-structure, materials and production-path question wearing a mechanical-engineering costume. The 2026 listing cycle put the debate on stage; the next two years of humanoid BOMs will render the verdict joint by joint — and the verdict, almost certainly, will be mixed.

Source: OFweek Robotics (October 6, 2026) — direct-drive vs. reducer route analysis; Securities Daily and China Rare Earth Group announcements (September 29–30, 2026) — strategic cooperation; Frost & Sullivan data cited in Benmo Technology listing materials — shipment volumes; Bernstein actuator cost analysis as cited by Chuangye Bang (March 2026) — 36% harmonic cost share. Company-sourced figures (torque density improvement) are identified as such in the text.