Harmonic vs Planetary Joint Modules for Humanoid Robots

Harmonic drives win where backlash and stiffness matter: wrists, forearms and any joint doing force control. Planetary drives win where torque per dollar, shock tolerance and speed matter: hips, knees and ankles. Most humanoids end up using both, because no single drive type covers the whole body well.

Why the drive type decides everything downstream

A joint module is a motor, a reducer, an encoder set and a controller in one housing. The motor and electronics are largely commodity. The reducer is not. It sets backlash, torsional stiffness, reflected inertia, backdrivability and the shock load the joint survives. Those five properties decide whether the joint can do force control at all, how much of the gearbox you have to model in software, and how often the robot comes back for service.

The honest question is not which reducer is better, but which failure mode this joint cannot afford. Precision joints cannot afford backlash; load-bearing joints cannot afford shock damage.

How the two drives work

Strain wave (harmonic) drives

A thin-walled flexspline with external teeth is deformed by an elliptical wave generator and meshes with a rigid circular spline that has two more teeth. The 2-tooth difference produces very high single-stage ratios, typically 50:1 to 160:1, with near-zero backlash because roughly 30 percent of the teeth are in mesh at any moment. The trade-off is a flexible intermediate element: torsional stiffness is finite, and the flexspline is the part that fatigues.

Planetary drives

Sun, planets and ring gear in a compact package. Load is shared across three or more planet gears, so torque density and shock tolerance are high and efficiency is several points better than a strain wave stage. Single-stage ratios are low, typically 3:1 to 10:1, so a useful robot joint needs two stages, which adds length, mass and accumulated backlash unless the unit is preloaded carefully.

Side-by-side, in the terms that matter

Read that table as constraints, not a ranking. A wrist needing 0.5 arcmin cannot use a two-stage planetary unit however cheap it is; a knee taking landing impacts will fatigue a flexspline long before a planetary set gives up.

Where harmonic modules win

• Wrist and forearm. Short lever arms, high ratio in one stage, and the joint must be light. A strain wave stage gets 100:1 into a housing you can hold in one hand.
• Anything doing force control. Backlash is a dead zone in the control loop. Below roughly 5 arcmin you can close a force loop on position; above it you spend your tuning budget chasing hysteresis.
• Shoulder pitch. A high ratio lets the motor run fast and small, which keeps copper loss and rotor inertia down.
• Joints where the reducer doubles as the safety element. A strain wave stage slips rather than shatters under moderate overload.

Where planetary modules win

• Hip and knee. Peak torque is large and transient; landing loads arrive as shocks. Planet sharing spreads that load and the gear set survives what a flexspline would not.
• Ankle roll and pitch on walking platforms. Short duty at high torque, frequent reversals, and a real need for efficiency so the battery is not spent on gear losses.
• Cost-sensitive platforms. Education, research copies and service robots where a 2-arcmin backlash is acceptable and 15 arcmin is not fatal.
• Joints that must be backdriven. Compliance control and physical human interaction are easier when the gearbox does not fight you.

Frame sizes and what they carry

Frame numbers are a packaging convention, not a performance promise. The values below are representative figures for our harmonic joint modules and are quoted to show the shape of the curve, not to replace a datasheet.

Figures are manufacturer-confirmed output values for the standard unit. A brake or an output torque sensor adds mass, and force-controlled variants are heavier still, so size against the datasheet of the exact ordering code.

Why most humanoids mix both

Walk the joint list and the split appears on its own: wrist, forearm and shoulder pitch want backlash near zero and mass near nothing, so they take strain wave stages; hip, knee and ankle want peak torque and shock survival, so they take planetary or quasi-direct-drive units.

That is why a single-technology body is rare outside renders. The practical question is which joints can share a frame size and a controller so the bill of materials stays manageable. For the arithmetic, the worked example in how to size joint actuators for a 7-DOF arm runs the gravity, inertia and duty terms joint by joint, and the harmonic drive actuator guide covers what changes when you push the ratio higher.

Three checks before you trust a comparison table

1. Is peak torque quoted at the output or the motor? A factor of the ratio sits between those two numbers, and datasheets are not consistent about which one they print.
2. Is backlash measured at the output with the brake engaged or free? Preload and brake drag both change the number by arcminutes.
3. Does the rated torque assume continuous or intermittent duty? A continuous rating at 100 percent duty is a different machine from an intermittent one, and the thermal time constant tells you which you bought.

Frequently asked

Can a planetary module achieve zero backlash?

Not in practice. Careful preloading gets a two-stage unit to roughly 3 to 5 arcmin when new; that figure grows with wear. Below 1 arcmin you are in strain wave or cycloidal territory.

Is a harmonic drive efficient enough for a walking robot?

At 70 to 85 percent it costs real battery on high-duty joints. That is why legs usually do not use it, while arms, which move less and need precision, do.

Do I need a torque sensor with either type?

Only if you need force control or collision detection at the output. A strain wave joint can estimate torque from motor current reasonably well; a planetary joint with more backlash estimates it less well.