How to Size Harmonic Drive Reducers for Robot Joints

Ask any humanoid robot engineer what the hardest component selection decision is, and they'll say the same thing: the joint reducer.

Not the motor. Not the encoder. Not the controller. The reducer.

The reducer (gearbox) is where torque, speed, precision, weight, cost, and lifetime all collide in one mechanical component. Get the torque rating wrong and you either shred the gearbox on the first jump or carry 30% more weight than you need to. There is no middle ground that doesn't hurt.

This guide breaks down how the industry actually sizes joint reducers: the calculation framework, the torque requirements by joint position, the reducer type selection logic, and the real-world derating factors that datasheets don't mention.

1. The Three Torque Regimes

Before talking numbers, it's important to understand that "how much torque" isn't one number. A humanoid robot joint operates in three distinct torque regimes, and the reducer must survive all three:

1.1 Continuous Torque (Thermal Limit)

This is the torque the joint can sustain indefinitely without overheating. It's governed by thermal equilibrium — the point where heat generation equals heat dissipation. For a sealed joint module with no active cooling, this is typically 40–60% of the peak rating on the datasheet.

1.2 Peak Torque (Dynamic Limit)

This is the maximum torque the joint experiences during dynamic motion — acceleration, deceleration, impact absorption, and gait transitions. It's transient (typically under 100 ms), but it's the torque that determines whether the reducer survives the robot's first week of testing.

Peak torque is where most sizing mistakes happen. Real peak torque includes:

• Static torque — gravity acting on the distal mass

• Inertial torque — accelerating the distal mass

• Impact torque — ground reaction forces during walking, landing, or collisions

• Back-EMF-limited motor torque — at high speeds, the motor can't deliver full peak

The DAF (Dynamic Amplification Factor) is the silent killer. A walking humanoid at 1.5 m/s needs a DAF of 2.0. A running or jumping humanoid needs 3.0–5.0. If you size for walking and then the control team decides the robot should jog, the reducers eat 2.5× the torque they were designed for.

1.3 Impact / Shock Torque (Survival Limit)

This is the instantaneous torque spike from ground impact, fall recovery, or unexpected collisions. It can reach 2–5× the rated peak torque and last 10–50 ms. The reducer doesn't need to function at this level — it needs to survive it without catastrophic failure (gear tooth fracture, spline stripping, bearing brinelling).

"Harmonic reducers are not recommended for core load-bearing joints (hip, knee). They are suited for shoulder, elbow, and wrist where impact loads are lower." — Industry Selection Guide

2. Torque Requirements by Joint Position

These are practical ranges compiled from industry sources, selection guides, and real robot teardowns. They represent requirements for a 60–80 kg full-size humanoid at walking speed (1.5 m/s):

2.1 Lower Body — The Power Zone

The hip is the monster. During stance phase, it supports the entire upper body mass (40–55 kg) at a moment arm of 0.30–0.40 m. That's 120–220 N·m of static torque alone. A 70 kg humanoid running needs hip pitch peak torque of 200–300 N·m. Jump? 400–600 N·m.

The knee is the second-most stressed joint. Unitree's G1 specs its knee at 90 N·m (standard) or 120 N·m (EDU) — but that's a 35 kg robot. Scale to 70 kg and the knee needs 150–200 N·m peak just for walking, 300+ for running.

The ankle takes the brunt of ground impact. It's the joint most likely to experience shock torque spikes — precisely why planetary reducers are preferred here.

2.2 Upper Body — The Precision Zone

Upper body joints prioritize precision and weight over raw torque. Harmonic reducers dominate here because they offer zero backlash (≤1 arcmin), high reduction ratios in a single stage, and compact form factors. The trade-off — poor shock tolerance — is acceptable because arms don't experience ground impact.

3. The Reducer Type Question

The standard framing of "harmonic vs. planetary" when discussing humanoid robot joints is incomplete. The actual landscape in 2026 is more nuanced:

3.1 Harmonic (Strain Wave Gear)

3.2 Planetary

Planetary reducers are making a comeback in humanoid lower body joints. The reason is simple: they survive shock loads. A planetary gear set distributes load across 3–5 planet gears, so a single impact doesn't concentrate stress. No fatigue accumulation from elastic deformation.

3.3 RV (Cycloidal)

RV reducers use a cycloidal cam mechanism that distributes load across many contact points simultaneously. They combine shock tolerance of planetary with the backlash performance of harmonic. Tesla's Optimus Gen 3 is reportedly shifting heavy-load joints to ultra-thin RV reducers that are 20% smaller and 30% stronger than the previous generation's harmonics.

3.4 Quasi-Direct Drive (QDD)

QDD is the fastest-growing segment. Instead of using a high-ratio reducer to multiply a small motor's torque, QDD uses a large, high-torque-density motor with a very low reduction ratio. The result: ±0.5% force control accuracy, sub-1ms response latency, and excellent backdrivability.

For a 35 kg humanoid like the G1, QDD makes sense — you can afford a slightly larger motor because the robot is already light. For a 70 kg humanoid, the motor size needed for QDD at the hip becomes impractical, and a planetary or RV reducer remains the better choice.

4. The Reduction Ratio Trap: Higher Isn't Better

One of the most counterintuitive findings: a higher reduction ratio can make a joint worse, not better.

Three things break when the ratio is pushed too high:

1. Stiffness disease — high-ratio joints can't passively comply. The robot can't absorb ground disturbances through joint compliance, requiring higher bandwidth control and more energy. It feels "robotic" rather than "human."

2. Flex spline deformation — ratios above 160:1 increase harmonic flex spline deformation by 20–30%, degrading transmission accuracy.

3. Back-EMF starvation — high ratios mean high motor speeds, where back-EMF reduces available torque. A 160:1 joint with a 5000 RPM motor outputs only 31 RPM — too slow for a hip that needs to swing 60° in 0.3 seconds.

5. The Datasheet Lie: Rated vs. Real-World Torque

Here's something no reducer datasheet will tell you: the torque rating on the spec sheet is measured under conditions the robot will never experience.

Reducer manufacturers rate their products at 20–25°C with continuous air cooling, lubricated at the correct input speed. Your robot's sealed joint housing will be 50–70°C with no convection and degrading grease.

6. How to Actually Size a Joint Reducer

Here's the practical workflow for teams building a humanoid robot:

1. Define the worst-case scenario — not nominal. For a knee: running at 2 m/s, landing from a 10 cm drop, carrying 5 kg.

2. Calculate swing-phase AND stance-phase torque — they're dramatically different. For legs, always use the higher value. A knee's swing torque might be 37 N·m; stance torque could be 165 N·m.

3. Apply safety factor and dynamic amplification — SF = 1.5 (conservative) to 2.0 (impact-prone); DAF = 2.0 (walking) to 5.0 (running/jumping).

4. Select reducer type — impact-heavy lower body → Planetary/RV; precision upper body → Harmonic; backdrivability → QDD.

5. Select reduction ratio — balance torque multiplication against output speed and compliance. Use the optimal range table above.

6. Verify derated continuous torque — multiply rated continuous torque by 0.57. If it's less than holding torque, go bigger.

7. Check fatigue life — for harmonics, verify L10 at design torque exceeds target lifetime. For leg joints, estimate fatigue consumption per impact cycle.

8. Verify motor can deliver — motor_peak_torque × reduction_ratio × efficiency ≥ T_design. If not, bigger motor or higher ratio — but remember the ratio trap.

7. What the Teardowns Show

Theory is one thing. Here's what teardown reports reveal about how real humanoid robots are actually sized:

Unitree G1 (35 kg)

Two-stage planetary reducers at ~20.58:1 in all leg joints. Knee peak: 90–120 N·m. This is a deliberately undersized design — the G1 is optimized for weight (35 kg) and dynamic agility (2 m/s+ running, backflips), not payload. The planetary reducers survive because the robot is light enough that ground impact forces are manageable.

Tesla Optimus Gen 2/3 (57–73 kg)

Hybrid approach — 14 harmonic reducers for rotary joints + 14 planetary roller screws for linear actuators. Gen 3 is reportedly shifting heavy-load joints to ultra-thin RV reducers. The roller screws cost $1,350–2,700 each — 14 of them is 19% of the total BOM.

CubeMars AKH70-48 (example module)

Peak torque 222 N·m at 1,396 g. Torque density: 159 N·m/kg. Planetary-reducer-based, designed for hip and knee. Features a 7 mm hollow shaft for cable routing — previously only available with harmonics. This is the trend: planetaries catching up on features while maintaining shock tolerance.

Key insight: When a manufacturer publishes 150+ N·m/kg torque density, they're engineering for weight. When they only publish peak torque without weight, they're hiding a heavy design.

Frequently Asked Questions