Every humanoid build reaches the same question: how much torque does this joint actually need? Size it too small and the gearbox fails in week one. Size it too large and you carry dead weight that cascades through the whole kinematic chain. This is the framework we walk customers through when they send us a load case.
Three torque regimes, three different numbers
“How much torque” is never one number. A joint has to survive three regimes, and the module you choose has to clear all three.
- Continuous (thermal) — torque the joint can hold indefinitely without overheating. In a sealed module with no active cooling this is typically 40–60% of the datasheet peak.
- Peak (dynamic) — the transient maximum during acceleration, deceleration and gait transitions, usually under 100 ms.
- Shock (survival) — instantaneous spikes from footfall, landing or an unexpected collision. These reach 2–5× peak and last 10–50 ms. The joint does not have to perform at this level, only survive it.
The usual sizing expression combines the first two: Tdesign = (Tstatic + Tinertial) × SF × DAF, where SF is a safety factor and DAF a dynamic amplification factor. DAF is where teams get hurt: 2.0 for a walking humanoid, 3.0–5.0 for running or jumping. Size for walking and then let the control team decide the robot should jog — every joint now eats 2.5× what it was designed for.
Torque by joint position
Practical ranges for a 60–80 kg full-size humanoid at 1.5 m/s:
| Joint | Rated (N·m) | Peak (N·m) | Reducer | Ratio |
|---|---|---|---|---|
| Hip (pitch / roll / yaw) | 80–200 | 240–600 | Planetary / RV | 50:1 – 80:1 |
| Knee (pitch) | 60–150 | 180–450 | Planetary | 50:1 – 80:1 |
| Ankle (pitch / roll) | 30–60 | 90–180 | Planetary | 50:1 – 80:1 |
| Waist (yaw) | 40–120 | 120–360 | Harmonic | 80:1 – 120:1 |
| Shoulder (3 axes) | 30–50 | 90–150 | Harmonic | 80:1 – 120:1 |
| Elbow (pitch) | 20–40 | 60–120 | Harmonic | 80:1 – 120:1 |
| Wrist (pitch / yaw) | 5–15 | 15–45 | Micro harmonic | 100:1 – 160:1 |
| Finger joints | 0.5–2 | 1.5–6 | Micro harmonic / planetary | 50:1 – 100:1 |
Ranges compiled from published selection guides and teardown reports. Treat them as a starting point, not a substitute for your own load case.
The pattern behind the table: planetary wherever the ground hits the robot, harmonic wherever precision and torque per gram matter. That split is the correct engineering answer, not a compromise — and we supply both, from Φ30 to Φ170 frames at 1.65 to 382 N·m rated.
The reduction ratio is not “more is better”
Three things break once you push the ratio past the useful band:
- Stiffness disease — a very high ratio joint cannot comply passively, so the controller has to absorb every ground disturbance. It costs bandwidth and energy, and the motion reads as rigid rather than human.
- Flexspline deformation — above roughly 160:1 in a harmonic stage, deformation grows 20–30% and transmission accuracy drops.
- Back-EMF starvation — higher ratios mean higher motor speed, where back-EMF eats available torque. A 160:1 joint on a 5000 rpm motor gives 31 rpm at the output: too slow for a hip that has to swing 60° in 0.3 s.
The number on the datasheet is a best case
Manufacturers rate modules at 20–25 °C with continuous air cooling and fresh grease. Your joint housing runs at 50–70 °C with no convection and nowhere for heat to go.
| Condition | Derating |
|---|---|
| Housing at 60 °C | ×0.75 |
| Sealed housing, no convection | ×0.85 |
| Grease degraded after ~5,000 h | ×0.90 |
| Combined real-world | ≈ ×0.57 |
A module rated 100 N·m continuous on paper delivers roughly 57 N·m inside a sealed humanoid joint after a few thousand hours. That gap is the entire reason experienced teams apply a 1.5–2.0× safety factor.
A seven-step selection workflow
- Define the worst case, not the nominal one. Running at 2 m/s, landing from a 10 cm drop, carrying 5 kg.
- Calculate swing and stance torque separately, and design to the higher. A knee might see 37 N·m in swing and 165 N·m in stance.
- Apply SF and DAF — SF 1.5 conservative to 2.0 impact-prone; DAF 2.0 walking to 5.0 running.
- Pick the reducer family — impact-heavy lower body gets planetary, precision upper body gets harmonic.
- Pick the ratio from the joint's useful band, not the highest one on the sheet.
- Verify the derated continuous torque — rated × 0.57 still has to exceed holding torque.
- Check the motor can deliver — motor peak × ratio × efficiency ≥ Tdesign.
Then sanity-check torque density in N·m/kg. When a maker publishes 150+ N·m/kg they are engineering for weight. When they publish peak torque with no weight figure at all, they are usually hiding a heavy design.
Send us the load case
Give us joint, torque, envelope, protocol and quantity, and we shortlist across every factory we buy from — not one catalogue. If you want to see the spread first, compare joint modules by torque band.
Specs quoted from manufacturer datasheets; final figures confirmed at quotation.