The Complete Robot Joint Module Selection Guide: Harmonic, Planetary, RV, Cycloidal & Quasi-Direct Drive — Vendor Research and Configurations by Robot Weight Class

Introduction: The Joint Module — a Robot's Muscle-Tendon System

If the motor is a robot's "muscle", the reducer is its "tendon" — it converts the motor's high speed into low-speed, high-torque output, and determines whether each joint can execute every motion precisely and stably. Pick the wrong transmission route and the entire robot loses, from the source, the ability to execute commands accurately.

In industrial robot cost structures, reducers account for roughly 30% — the single largest core component cost. Humanoid robots need far more reducers than traditional industrial robots (which use 2-6 units each): a humanoid with 28-42 rotary joints needs nearly as many reducers as motors. This makes joint-module selection both a performance problem and a cost problem — if per-joint transmission cost cannot be pushed below ¥1,500, total transmission cost exceeds ¥60,000-80,000, i.e. 30%-40% of BOM.

This article draws on 200+ joint-module/reducer entries in the RobotWorld knowledge base, product research on leading joint-module and reducer vendors (Harmonic Drive, Nabtesco, Leaderdrive, Laifual, Leadshine, EYOBOT, Juxie Drive, ZeroErr, JIECANG, CubeMars, ENCOS, DaMiao and more), item-by-item verification of the supply chains behind Tesla, Unitree, UBTECH, AgiBot and Fourier, plus industry research and public sources (Xueqiu/Eastmoney supply-chain teardowns, GGII, CSDN teardowns) to answer one question systematically: how should robot joint modules actually be selected? Where do harmonic, planetary, RV and cycloidal each belong — and how should different-sized robots be configured?

The Four Transmission Routes at a Glance

There are four mainstream transmission routes for robot joints today, plus a rapidly rising quasi-direct-drive (QDD) approach. Overview table first:

Route	Single-stage ratio	Backlash	Efficiency	Life	Cost tier	Typical placement
Harmonic (strain-wave)	50-320	≤1 arcmin (precision 10-20 arcsec)	70%-90%	6k-10k h	High	Wrist, forearm, hand, ankle
Planetary	3-10 (multi-stage to 100)	3-8 arcmin (precision ≤3)	95%-97%	>20k h	Lowest (single stage: a few hundred CNY)	Hip/knee low-ratio QDD joints
RV reducer	30-260	≤1 arcmin	85%-92%	20k-30k h	Highest	Base, upper arm, shoulder heavy loads
Cycloidal (pure)	11-87	1-3 arcmin	90%-94%	High	Mid-high (domestic ~50% of Japanese price)	High-torque compact joints, base
Quasi-direct drive (QDD)	4-10	Encoder-dependent	>95%	High	Low	Ankle, hip, knee force-control joints

The industry rule of thumb: harmonic for precision light loads, planetary for mid-load dynamic duty, RV for heavy loads. But a humanoid's joint torque demand spans from 0.5 N·m in dexterous-hand fingers to ~200 N·m impact at the hip — a 40x range. No single reducer can cover it all; real designs must mix routes joint by joint.

Cross-Route Performance Comparison

Putting the four routes on one radar chart makes each one's strengths and weaknesses immediately visible:

Dimension	Harmonic	Planetary	RV	Note
Backlash	<1 arcmin	3-5 arcmin	1-2 arcmin	Harmonic's zero backlash is principled (multi-tooth meshing of flexspline)
Typical ratio	50-160	3-10/stage	30-200	Harmonic reaches 100+ single-stage; planetary needs multi-stage trains
Torsional stiffness	~1e4 N·m/rad	~3e4	~5e4	RV's dual-supported structure is stiffest
Precision life	8k-10k h	15k-20k h	20k-30k h	Flexspline fatigue causes harmonic precision to decay over time
Cost index	1.5	1.0	2.0	Single-stage planetary = baseline 1.0

Data sources: GB/T 30819-2014 (harmonic gear reducers for robots), GB/T 37718-2019 (precision cycloidal-pin reducers for robots), GB/T 41400-2022 (precision planetary gear reducers for robots) plus public industry comparison tests. The three national standards cover exactly the harmonic, cycloidal/RV and planetary routes — cite the corresponding standard as your acceptance-test basis.

Special Constraints of Humanoid Robots

Industrial-robot selection only weighs the torque-precision-life triangle. Humanoids add four harsher constraints:

1. Joint count x5-7, cost must halve

A six-axis industrial robot has 6 joints and uses 2-6 reducers; a humanoid has 28-42 rotary joints. Reusing industrial-grade harmonic drives (¥3,000+ each) would push total transmission cost past ¥100,000. The industry target is ≤¥1,500 per joint; aggressive designs (planetary/cycloidal-dominant legs) push below ¥800:

2. Every gram counts

Mainstream humanoids weigh 55-80 kg. Per MIIT Fifth Electronics Research Institute estimates, every 1 kg of weight saved yields roughly 3%-5% more runtime. Joint modules are the most widely distributed mass on a robot — saving 100 g at each of 40 joints buys 4 kg of runtime dividend. This is also why hollow-shaft designs (cables routed through the axis) are popular: no external cable trays, indirectly saving structural mass.

3. Impact loads and force control

During walking, running and landing, hip/knee joints absorb ~200 N·m-class transient impacts; the ankle additionally needs good back-drivability for force-controlled landing cushioning. Harmonic flexsplines age dramatically faster under repeated impact — the core reason leg joints avoid going all-harmonic.

4. Precision requirements can actually relax

A counterintuitive conclusion: if assembly and compensation algorithms guarantee ±1 mm end-effector accuracy, per-joint backlash can relax from industrial-grade <1 arcmin to ~3 arcmin. Humanoids therefore need not pay the industrial precision premium at every joint — this is what makes planetary's "good-enough" precision viable.

The Four Routes in Depth

Harmonic: the dexterous hand of precision light loads

A harmonic reducer has three core parts — wave generator, flexspline and circular spline — transmitting motion through elastic deformation of the flexspline. Single-stage ratios of 50-320, backlash within 10 arcseconds, small, light, zero-backlash: these traits make it the de-facto standard for precision light-load joints.

Fatal weakness: the flexspline flexes continuously and fatigues; life shortens sharply under sustained heavy or impact loads, and motion precision decays with hours of use. Efficiency also drops at high input speeds (into the low 70s %), with concentrated heat.

Best fit: forearms, wrists, hands and ankles — joints demanding precision and agility under light loads. Market: Japan's Harmonic Drive holds roughly 50%-60% of global share (some estimates run higher); domestically Leaderdrive leads — 2025 harmonic sales of 425,000 units, +72.5% YoY (~246k in 2024), holding 12%-15% global share — followed by Laifual and Han's. Leaderdrive's LCS/LCSG cup series covers 4-969 N·m rated torque at ratios 30-160 with ≤10 arcsec backlash; Laifual's FH hollow-shaft harmonic targets humanoid cable-routing scenarios.

Planetary: the pragmatic choice for mid-load dynamics

A planetary gearbox (sun, planets, ring) delivers 3-10:1 per stage (multi-stage trains reach 100:1), 95%+ efficiency, compact structure, strong load capacity — and the best cost of all routes, with single-stage units priced in the hundreds of CNY.

Limitations: small per-stage ratio, standard backlash 3-8 arcmin, precision degrades under impact. But precision-grade planetary (ground gears + preload backlash elimination) reaches ≤3 arcmin at 2-3x the price.

Best fit: cost-sensitive, moderate-precision mid-load oscillating positions. Humanoid legs are adopting planetary at scale — the hip/knee joints of Unitree and AgiBot platforms, and mainstream modules like the CubeMars AK series and ENCOS EC-A series, lean heavily on 9:1-48:1 planetary. Germany's Neugart (PLE economy / PSNpro precision) remains the international benchmark.

RV: the load-bearing powerhouse

The RV reducer (planetary front stage + cycloidal-pin rear stage) delivers 85%-90% efficiency and the highest fatigue strength, stiffness and life of the four routes, with precision that — unlike harmonic — does not decay significantly over time.

Weaknesses: heavy, bulky, complex, and the most expensive (cost index ~2.0). The classic six-axis industrial configuration is "2 RV + 4 harmonic" — RV on base and upper arm, harmonic on forearm and wrist. Market: Nabtesco holds ~52%-60% global share; its China share fell from 52.3% (2024) to 42.7% (2025), while Shuanghuan — China's #1 domestic RV maker at ~12.5%-15% — plus Qinchuan, Huandong and Zhongda Leader price RV at 60%-70% of Japanese levels and are substituting fast. RV is still rare in humanoids, but lightweight micro-RV programs are underway.

Cycloidal: the underrated torque-density champion

Pure cycloidal-pin reducers share DNA with RV (RV's rear stage is cycloidal-pin) as single-stage standalone products: ratios 11-87, backlash 1-3 arcmin, efficiency 90%-94%. Their core strengths are extreme torque density and overload capability — simultaneous multi-tooth meshing tolerates 5x+ transient overload; Sumitomo's Fine Cyclo A series spans 785-24,000 N·m rated torque, the ceiling for heavy loads.

Slovakia's SPINEA (Timken) integrates cycloidal gearing with crossed-roller bearings in the TwinSpin T series: 37-2,940 N·m rated, ratios 33-191, emergency-stop torque up to 5x rated; its DriveSpin DSH hollow-shaft cycloidal actuators (18-420 N·m rated, 8-40 mm bore) directly target large humanoid joints. Domestically, Zhongda Leader and Jiangsu Tailong price cycloidal at ~50% of Japanese levels. Cycloidal's drawbacks: precision below harmonic and high manufacturing barriers — it is a route "ramping up" in humanoids today.

Quasi-Direct Drive: the Paradigm Shift Humanoids Brought

QDD is the most important transmission trend in humanoid robots: low-ratio (4-10:1) planetary/cycloidal plus a high-torque-density motor, trading some torque multiplication for three dividends:

• Back-drivability: external force reflects to the motor side attenuated only by 1/i, while reflected friction scales with i² — dropping the ratio from 100 to 10 cuts reflected friction ~100x, making the joint "transparent" and naturally suited to force control (ankle cushioning, human-robot interaction compliance);
• Impact tolerance: no elastic weak part like a flexspline; impacts are absorbed directly by gears and rotor inertia;
• Low cost, high efficiency: single-stage planetary costs hundreds of CNY at 95%+ efficiency.

QDD's costs: continuous motor torque is limited to ~30%-40% of peak (thermal limit), and friction compensation / disturbance observers are required to cancel residual friction. Xiaomi CyberOne, Unitree, the MIT Cheetah lineage and RobStride all use this route for legs. RobStride pushes QDD module cost-performance to the extreme: RobStride 04 (9:1, 120 N·m peak, 1.42 kg) at ¥1,199 and RobStride 05 (7.75:1, 191 g) at just ¥499.

Vendor Landscape

By entry count in the RobotWorld knowledge base (a proxy for product-line breadth and documentation completeness), domestic module vendors are already more active than Japanese ones:

Vendor	Entries	Main route	Representative products / notes
Harmonic Drive	19	Harmonic components	CSF/CSG/FB component sets, global harmonic leader
Kaiserdrive (Kaixuan)	16	Harmonic modules	Full KGU-08~40 line, 12-762 N·m peak, zero backlash
Juxie Drive / EYOBOT	15 / 15	Harmonic + planetary	EYOBOT covers both PHA/PHU harmonic and RP planetary lines
DaMiao Technology	14	Harmonic + planetary modules	DM-JH harmonic 10.5-24 N·m; J6248P 97 N·m peak
Leadshine / ENCOS	14 / 10	Dual harmonic + planetary	Leadshine HJ2 harmonic 11-194 N·m; ENCOS EC-A full spectrum
CubeMars (Sanrui)	9	Planetary QDD	AK series 22-222 N·m; AKH70-48 hollow-shaft ¥4,888
RobStride / HCFA	9 / 9	QDD / harmonic + planetary	RobStride ¥499-1,199; HCFA Hu-MRS harmonic at 10 arcsec
Leaderdrive (Lüdi)	8	Harmonic reducers	LCS/LCSG/LHD; China's #1 harmonic, 425k units sold in 2025 (+72.5%)
ESTUN Codroid / Windouble / Tongchuan	6 / 5 / 5	Harmonic modules	Codroid SA series with integrated torque sensor; Windouble KSG up to 1,600 N·m
SPINEA / Laifual / Newgear	5 / 3 / 3	Cycloidal / harmonic	SPINEA TwinSpin cycloidal benchmark; Laifual hollow-shaft harmonic + dual-voltage modules

Also catalogued: Shuanghuan and Qinchuan (RV), Zhongda Leader (harmonic/planetary/cycloidal in parallel — ZD-JM01/02/03), Nabtesco (RV-E/RV-C), Sumitomo (Fine Cyclo), Neugart (planetary), HIWIN (DATORKER harmonic), INNFOS (SCA smart joints), RealMan (WHG modules), MagicLab (H70 harmonic, 96 N·m / 900 g). The sector shows a clear division of labor: Japanese vendors defend reducers, Chinese vendors attack integrated modules.

Integrated joint-module vendors: who can you actually trust

Buying a standalone reducer is only half the story — humanoid mass production runs on integrated modules (motor + transmission + dual encoders + driver). A 2026 rotary-module brand ranking (Daishi Caijing / industry media) is a useful reference, combined with our supply-chain verification:

Vendor	Route	Representative products	Mass-production track record (verified)
Leadshine	Frameless motors + harmonic/planetary modules	HJ2 harmonic modules 11-194 N·m; FM frameless torque motors	AgiBot (70%+ frameless-motor share, 2026 "Outstanding Supplier", 10k+/month module deliveries), Unitree (50k+ frameless delivered in 2024), UBTECH (15 force-control joints on Walker X, 20k/month orders), Fourier/Galaxea/Dreame; Optimus frameless second-source (via Sanhua/Tuopu)
EYOBOT (Yiyou)	Integrated harmonic joints + planetary	PHA/PHU harmonic, RP planetary, high-torque-density integrated joints	AgiBot A2 series mass-production designated supplier; Huashu, Kailda
Juxie Drive	Self-developed harmonic modules	Humanoid full joint-module kits	Knowledge-base entry count in the top tier; active in humanoid integrated solutions
ZeroErr	eRob rotary joint modules	Inline-I / angle-T types, full torque spectrum	Major supplier to collaborative and multi-joint robots (Shenzhen)
Taihu Robotics	Professional joint modules	Humanoid-oriented joint modules	Ranked into the 2026 rotary-module TOP10
JIECANG	Harmonic / planetary / cycloidal, three series	JCRM50-70 (peak 30-180 N·m), JCPM52-124 (rated 4-100 N·m), dual-steel-wheel harmonic; 15 N·m/kg torque density, ±0.01° accuracy	Batch shipments to humanoid joints already running; listed company, 100% self-developed core parts, cycloidal -35% weight / -50% cost
Tongchuan Technology	Harmonic / integrated joints	Harmocore lightweight harmonic joints, TRM series 46-672 N·m	Full-spectrum integrated joints, KB-catalogued
ENCOS (Yinkesi)	Joint actuators/modules	EC-A full spectrum (including hollow-shaft variants)	KB-catalogued; popular in research and education platforms
Leaderdrive (Lüdi)	Harmonic reducer → rotary actuators	LCS/LCSG/LHD reducers, PHT integrated joints, rotary-executor JV with Sanhua	2025 harmonic sales 425k units (+72.5%), 12%-15% global share; Tesla Optimus rotary-joint harmonic supplier, UBTECH Walker S2 primary harmonic supplier (Nantong base, 1.2M-set capacity)
Laifual	Harmonic + modules	FH hollow-shaft harmonic, L52I-L170I modules	Core transmission supplier to multiple module makers; extending downstream into modules
Harmonic Drive	Harmonic components + FHA-C actuators	CSF/CSG components, FHA-C rotary actuators	~50-60% global harmonic share; the premium benchmark

Module-selection advice: prefer vendors with both self-developed core components and mass-production wins at head robot companies — the former determines cost and delivery stability, the latter is the hardest reliability endorsement. Leadshine and EYOBOT already hold mass-production designations at AgiBot and UBTECH; JIECANG is the listed player covering all three routes (harmonic/planetary/cycloidal); Leaderdrive is the only reducer giant combining harmonic scale with module extension. For pure procurement, qualify 2-3 module vendors and test them in parallel.

What the head robot companies actually use

Whose joints do the leaders really buy? A supply-chain map after item-by-item verification (sources: Xueqiu/Eastmoney supply-chain teardowns, company disclosures and third-party reports):

Robot company	Joint architecture	Verified supply chain
Tesla Optimus	28 body joints = 14 rotary (frameless torque motor + harmonic reducer + output encoder/torque sensor, in 20/110/180 N·m tiers) + 14 linear (frameless motor + planetary roller screw), converged to 6 actuator types (3 rotary + 3 linear)	Leaderdrive (rotary-joint harmonics), Moons' / Zhaowei (dexterous-hand micro drives), Tuopu / Sanhua (actuator Tier-1 integrators), Fulin Precision (electric-joint co-development)
UBTECH Walker S2	Servo drivers 100% self-developed (170+ patents, core parts >90% localized); Walker's 36 servo joints mostly harmonic, Walker X/S's 41 joints with self-developed harmonic/planetary integrated drivers	Leaderdrive (main harmonic supplier, Nantong base), ULI (6-axis force sensors), Keli Sensing (torque sensors); 2025 delivered 1,079 units, avg ¥760k, 54.6% gross margin — proof that in-house joints pay
Unitree	Fully self-developed joints: motors, reducers, encoders, drivers ("upstream of joint motors there are only copper-wire suppliers"); G1 dual-encoder two-stage planetary modules (total ratio ~20.6), 140 N·m peak instant torque	Leadshine (servo/frameless, 50k+ frameless in 2024), Moons' (control motor modules), Jiangsu Leili ("brushless motor + gearbox" planetary actuators), Zhongda Leader (precision reducers)
AgiBot	PowerFlow self-developed joint motors + integrated harmonic joints + high-transparency planetary (ratio <10); A2 with 49 DoF total	EYOBOT (A2 mass-production designated integrated joints), Leadshine (70%+ frameless share, 10k+/month modules), Fulin Precision (harmonic/joint-assembly co-development)
Fourier	GR-1: 32 fully self-developed FSA integrated actuators — harmonic upper body, planetary hips	Full actuator stack in-house; among the earliest Chinese vendors to open-source the platform (Fourier N1, 23 DoF)

Three observations: ① head OEMs universally self-develop or deeply co-develop their joints — pure off-the-shelf buying is now rare in mass-production humanoids, so module vendors enter as "designated suppliers / co-development partners"; ② harmonic remains dominant in rotary joints (all 14 Optimus rotary joints use harmonic), while planetary / roller screws take linear and high-impact joints; ③ a "head-customer designation" is the most credible endorsement a module vendor can hold — Leadshine, EYOBOT and Fulin Precision already ship at volume into AgiBot/UBTECH/Unitree, so shortlist them first.

Joint Module Quick Reference: by Torque Tier

Sorting fully-specified modules from the knowledge base by peak torque shows domestic modules already span the full spectrum from dexterous hands to large joints:

({
  "_height": "420px",
  "backgroundColor": "transparent",
  "grid": { "left": 170, "right": 90, "top": 30, "bottom": 45 },
  "tooltip": {
    "trigger": "axis",
    "axisPointer": { "type": "shadow" },
    "backgroundColor": "rgba(15,23,42,0.95)",
    "borderColor": "#3ddc97",
    "borderWidth": 1,
    "textStyle": { "color": "#e2e8f0" },
    "formatter": function(params) {
      var p = params[0];
      var info = [
        "Harmonic 101:1 · 207 g",
        "Planetary 7.75:1 · 191 g · ¥499",
        "Harmonic 101:1 · 345 g",
        "Planetary 7.75:1 · 405 g · ¥699",
        "Harmonic 51:1 · 985 g",
        "Planetary 9:1 · 540 g",
        "Planetary 9:1 · 940 g",
        "Harmonic 101:1 · 900 g",
        "Planetary 39:1 · 750 g",
        "Harmonic · 900 g",
        "Planetary 9:1 · 1.42 kg · ¥1199",
        "Harmonic 101:1 · 2.0 kg · 10 arcsec",
        "Planetary 48:1 hollow-shaft · 1.4 kg · ¥4888",
        "Harmonic 120:1 · 5.5 kg",
        "Harmonic 121:1 · 7.9 kg",
        "Harmonic 160:1 · 22 kg"
      ];
      return p.name + "
Peak torque: " + p.value + " N·m
" + info[p.dataIndex];
    }
  },
  "xAxis": {
    "type": "log",
    "min": 1,
    "max": 3000,
    "name": "Peak torque (N·m, log axis)",
    "nameTextStyle": { "color": "#64748b" },
    "axisLabel": { "color": "#64748b" },
    "splitLine": { "lineStyle": { "color": "rgba(148,163,184,0.12)" } }
  },
  "yAxis": {
    "type": "category",
    "data": [
      "EYOBOT PHA08L-101",
      "RobStride 05",
      "DaMiao DM-JH11-101",
      "RobStride 02",
      "DaMiao DM-JH17-51",
      "CubeMars AK70-9",
      "CubeMars AK10-9",
      "EYOBOT PHA17H-101",
      "CubeMars AK60-39",
      "MagicLab H70",
      "RobStride 04",
      "HCFA Hu-MRS220N",
      "CubeMars AKH70-48",
      "Windouble KSG-20-120",
      "Kaiserdrive KGU-40D",
      "Windouble KSG-32-160"
    ],
    "axisLabel": { "color": "#334155", "fontSize": 12 },
    "axisTick": { "show": false }
  },
  "series": [
    {
      "type": "bar",
      "barWidth": 13,
      "data": [
        { "value": 3.3, "itemStyle": { "color": "#e4572e" } },
        { "value": 5.5, "itemStyle": { "color": "#2f6df6" } },
        { "value": 10.5, "itemStyle": { "color": "#e4572e" } },
        { "value": 17, "itemStyle": { "color": "#2f6df6" } },
        { "value": 24, "itemStyle": { "color": "#e4572e" } },
        { "value": 29.2, "itemStyle": { "color": "#2f6df6" } },
        { "value": 53, "itemStyle": { "color": "#2f6df6" } },
        { "value": 66, "itemStyle": { "color": "#e4572e" } },
        { "value": 72, "itemStyle": { "color": "#2f6df6" } },
        { "value": 96, "itemStyle": { "color": "#e4572e" } },
        { "value": 120, "itemStyle": { "color": "#2f6df6" } },
        { "value": 220, "itemStyle": { "color": "#e4572e" } },
        { "value": 222, "itemStyle": { "color": "#2f6df6" } },
        { "value": 360, "itemStyle": { "color": "#e4572e" } },
        { "value": 762, "itemStyle": { "color": "#e4572e" } },
        { "value": 1600, "itemStyle": { "color": "#e4572e" } }
      ],
      "label": { "show": true, "position": "right", "color": "#334155", "fontSize": 11, "formatter": "{c} N·m" }
    }
  ]
})

A few anchor products worth noting:

Module	Transmission	Rated / peak torque	Weight	Highlight
CubeMars AKH70-48	Planetary 48:1	74 / 222 N·m	1,396 g	159 N·m/kg torque density, 7 mm hollow bore, ¥4,888
ENCOS EC-A9025-P2	Two-stage planetary 23.6:1	100 / 320 N·m (est.)	1,690 g	Hollow cable routing, built for high-load hips
ESTUN Codroid SA20	Harmonic 101:1	52 N·m (rated)	—	Integrated torque sensor, force-control friendly
HCFA Hu-MRS220N	Harmonic 101:1	120 / 220 N·m	2.0 kg	10 arcsec backlash, high-torque harmonic
Windouble KSG-32-160	Harmonic 160:1	800 / 1,600 N·m	22 kg	Domestic high-torque harmonic ceiling
Tongchuan TRM series	Harmonic modules	46-672 N·m	1.6-8.2 kg	Full integrated-joint spectrum
Laifual L52I-L170I	Harmonic 51-161:1	11.5-1,530 N·m	0.55-9.48 kg	DC low-voltage joint module line; one of the top harmonic shippers in China by 2025 shipment volume

Sizing by Robot Weight Class

Different robot sizes face entirely different torque demands, weight budgets and cost sensitivities. Three weight classes, three configurations:

Class	Hip / knee	Ankle	Shoulder / elbow	Wrist / hand	Reference platforms
Small ≤25 kg (education, research, light service)	Single-stage planetary QDD 9:1 (RobStride 02/03, AK70-9; ¥500-1,000 each)	Same as hip/knee or harmonic size 14	Harmonic size 14/17 (EC-A3814 434 g, PHA08L 207 g)	Micro harmonic 11/14 or micro planetary	RobStride education kits, Unitree Go derivatives
Medium 30-55 kg (general-purpose humanoid mainstream)	Planetary 9-48:1 (AK60-39, AKH70-48, EC-A6416) or harmonic size 20/25	Harmonic size 20/25 (EC-A6013 130 N·m)	Harmonic size 17/20 (EC-A5013 90 N·m, DM-JH17)	Harmonic size 14 + dexterous-hand micro modules (0.5-5 N·m)	Unitree G1 (23-43 DoF, from ~$16k)
Large 60-100 kg (high-performance humanoids, heavy duty)	Two-stage planetary 23-25:1 (EC-A9025 320 N·m) or large harmonic/micro RV (KSG-20 360 N·m, TRM 672 N·m)	Harmonic size 25/32 (KSG-17 180 N·m) or cycloidal DSH	Harmonic size 25/32 (KGU-25 198 N·m, Hu-MRS220N)	Harmonic size 17/20	Unitree H1, industrial-grade humanoids

Three rules of thumb:

• The heavier the robot, the more its legs should avoid pure harmonic: hips/knees of 60 kg+ humanoids see impacts near 200 N·m and flexspline fatigue risk spikes — planetary/cycloidal/large harmonic is the safer choice;
• The lighter the robot, the more it can go all-harmonic: under 25 kg, impact loads are small, harmonic's precision and lightness shine, and cost stays acceptable;
• Upper limbs always lean harmonic: regardless of size, shoulder/elbow/wrist loads sit inside harmonic's comfort zone; only heavy-duty industrial arms need RV or large cycloidal at shoulder/base.

Industry Practice: How the Leaders Configure

Six-axis industrial robots: 2 RV + 4 harmonic

The classic recipe persists: RV on base and upper-arm axes (heavy load, high stiffness), harmonic on the three forearm/wrist axes (light, small, precise). The logic: stiffness for heavy loads, precision for light loads.

Tesla Optimus: rotary harmonic + linear roller screws

Teardowns converge on a clear architecture: 28 body joints = 14 rotary (frameless torque motor + harmonic reducer + output encoder/torque sensor, in 20/110/180 N·m tiers) + 14 linear (frameless motor + planetary roller screw) — the elbow/knee/ankle flexion positions use linear roller-screw actuators rather than planetary gear reducers — converged to 6 actuator types (3 rotary + 3 linear). Harmonic on the rotary side, roller screws on the linear side, frameless motors and structure self-supplied: the textbook case of "designing transmission for mass production."

Unitree: fully in-house, planetary-dominant

Unitree is the most thorough self-developer: joint motors, reducers, encoders and drivers are all built in-house (official line: "upstream of joint motors there are only raw materials like copper wire"). G1 uses dual-encoder two-stage planetary joint modules (total ratio ~20.6) with 140 N·m peak instant torque — harmonic upper limbs, planetary-dominant legs — holding the starting price at ~$16k through a simplified actuator architecture; H1, the performance platform, emphasizes joint torque density and positional accuracy. External supply concentrates at the component level: Leadshine (servo/frameless motors, 50k+ frameless delivered in 2024), Moons' (control motor modules), Jiangsu Leili ("brushless motor + gearbox" planetary actuators), Zhongda Leader (precision reducers).

The emerging "hybrid template"

Across teardowns, humanoid transmission is converging on a template: harmonic upper limbs, small RV or large harmonic at shoulder/elbow, planetary or cycloidal at hip/knee, harmonic at the ankle (small/medium robots) or planetary QDD (large robots).

Per-Joint Selection Decision Tree

flowchart TD
    A["Three questions per joint:
How much torque? Force control needed? What budget?"] --> B{"Peak torque > 150 N·m?"}
    B -->|"Yes"| C{"Heavy impact (walking/jumping)?"}
    C -->|"Yes"| D["Two-stage planetary 20-50:1
or cycloidal / micro RV"]
    C -->|"No"| E["Large harmonic size 25-40
(KSG/TRM class)"]
    B -->|"No"| F{"Force control / back-drivability needed?"}
    F -->|"Yes"| G["QDD: single-stage planetary 4-12:1
+ friction compensation"]
    F -->|"No"| H{"Torque > 40 N·m?"}
    H -->|"Yes"| I["Harmonic size 20/25
or planetary 30-50:1"]
    H -->|"No"| J{"Precision < 1 arcmin required?"}
    J -->|"Yes"| K["Harmonic size 11-17"]
    J -->|"No"| L["Precision planetary vs harmonic,
decide by cost"]

Selection Pitfalls: A Practical Guide

From public cases and vendor interviews, the six most common traps:

• Pitfall 1: Matching on motor power alone, ignoring load inertia and peak torque.Under fast start-stop, inertia mismatch causes oscillation and overcurrent. Compute peak torque = load torque + inertial acceleration torque, then keep 1.5-2x margin.
• Pitfall 2: Confusing rated with peak, and peak with continuous.Peak torque typically holds for only hundreds of milliseconds to seconds; continuous torque is ~30%-40% of peak (especially for QDD modules). Always verify thermal behavior against the continuous duty cycle.
• Pitfall 3: Blindly choosing harmonic for impact duty.Flexspline fatigue life is extremely impact-sensitive; for hip/knee/ankle either switch routes or oversize the harmonic by 2+ size classes.
• Pitfall 4: Ignoring backlash propagation.Backlash in serial joints accumulates and amplifies at the end effector — but if ±1 mm end accuracy suffices, relaxing to ~3 arcmin and choosing precision planetary saves 40%+ cost.
• Pitfall 5: Comparing reducer unit prices, ignoring integration cost.Discrete builds (motor + reducer + encoder + driver bought separately) look cheaper, but assembly, calibration and cabling engineering can exceed the module premium. For humanoid mass production, integrated joint modules (motor + transmission + dual encoders + FOC drive + CAN) are virtually inevitable.
• Pitfall 6: Forgetting cable routing.Cable management across dozens of joints is a hidden nightmare; hollow shafts (harmonic hollow-bore versions, AKH70, EC-A*H, DriveSpin DSH) are worth paying a premium for.

Final Conclusions and Implementation Path

Compressing the article into an actionable selection path:

Step	Action	Output
1. Define the spectrum	List every joint's peak/continuous torque, speed and impact class	Joint torque spectrum (cf. Fig. 1)
2. Assign routes	Decision-tree per joint: harmonic upper limbs, planetary/cycloidal legs, QDD for force-control joints	Route assignment table
3. Match modules	Pick candidates by torque tier from the knowledge base (quick-ref table above); prefer integrated + hollow-shaft	Candidate module list
4. Verify	Continuous-torque thermals, inertia ratio (<3:1 force control, <10:1 position control), backlash accumulation, life (harmonic 6k-10k h)	Verification report
5. Negotiate	Target ≤¥1,500/joint for humanoid mass production; ≤¥3,000 for industrial reuse lines	BOM cost model
6. Dual-source	Qualify ≥2 harmonic vendors (e.g., Leaderdrive + Laifual/HD); domestic planetary supply is mature and multi-source	Supply-chain resilience plan

One-sentence summary: there is no best reducer — only the right reducer in the right place. Give upper limbs to harmonic, legs to planetary and cycloidal, force-control joints to low-ratio QDD, and cost to domestic integrated modules.

Uncertainties and Information Gaps

Honest limitations:

• Joint torque figures are engineering estimates for 55-80 kg humanoids; specific machines can vary ±30%. Final selection must rest on whole-body dynamics simulation.
• Market-share numbers (HD ~50-60%, Nabtesco ~52-60%, Leaderdrive 425k units in 2025) come from company reports and industry research; measurement bases (units vs revenue, domestic vs global) are not fully consistent.
• Domestic module "rated/peak" labeling is not standardized (some values are vendor estimates, e.g., certain ENCOS models); cross-vendor comparisons need caution — trust measurements.
• Tesla Optimus transmission details come from third-party teardowns and reports, not official sources; Xiaomi and Figure have not published complete transmission BOMs.
• Head-company supply-chain relationships (Leadshine×AgiBot/UBTECH, EYOBOT×AgiBot, Leaderdrive×UBTECH/Tesla, etc.) come from Xueqiu/Eastmoney supply-chain teardowns and media reports — partly industry-channel research, not company filings.
• Cycloidal adoption in humanoids is still early; SPINEA/Sumitomo humanoid deployments are limited. This article's cycloidal outlook is mildly optimistic — keep tracking.
• Knowledge-base entry counts reflect catalog richness, not market share.