BLOGS

Ask for a robot joint's torque, its speed, its gear ratio, even its torsional stiffness, and any datasheet will answer. Ask for its first resonance — the frequency where the whole drivetrain starts to ring — and you will usually get silence. Yet this uncharted number decides how aggressively you can command torque, how hard your controller can push gains, and whether the joint hums, buzzes, or howls under load. This piece maps the resonance problem from first principles to mitigation, with the real numbers engineers work from.

Two Inertias and a Spring: Where Resonance Comes From

Every geared robot joint is, mechanically, two inertias coupled by a spring. On the motor side sits the rotor inertia. On the load side sits the link, reflected through the gear ratio. Between them lives the transmission's torsional stiffness — in an integrated joint, that spring is overwhelmingly the gear itself, not the shafts or couplings around it.

Model it as the classic two-inertia system and two frequencies fall out: a resonance pole, where motor and load swing against each other, and an anti-resonance zero, where they cancel. The pole sits at ω = √(K·(1/Jm + 1/JL)), the zero at √(K/JL), with K the drivetrain stiffness, Jm the motor-side inertia and JL the load-side inertia seen through the gearing. The pole is always above the zero. Neither appears on a datasheet, but both are measurable in any frequency-response test of the drivetrain.

These are not exotic numbers. A 2021 experimental study of an industrial servo drivetrain modeled exactly this way measured anti-resonance frequencies of 44.8 to 80.4 Hz and resonance frequencies of 51.8 to 93.1 Hz across its operating configurations, with simulation-to-experiment errors under 10 Hz (Zhang et al., Mechanical Sciences 12:305, 2021, doi:10.5194/ms-12-305-2021). A harmonic-drive test bench studied at McGill showed the same structure at lower frequencies still — resonance features near 200 and 400 rad/s, roughly 32 and 64 Hz (Taghirad, "Robust Torque Control of Harmonic Drive Systems," McGill University, 1997). Somewhere between a few tens and about a hundred hertz is where a real joint's first mode typically lives — squarely inside the bandwidth region where a modern servo is asked to work.

The rule that follows from the two-inertia picture: the controller's effective bandwidth must stay comfortably below the first resonance. Push gain past the pole and the loop that was supposed to regulate torque becomes the excitation source. Exactly how far below is a tuning judgment — vendors differ — but "well under the first mode" is not negotiable, and the notch filter is the industry's standard patch when it is (Galil Application Note #3502, "Using Notch Filters to Mitigate Resonance").

The Strain Wave's Hidden Spring: Stiffness That Changes With Load

The two-inertia model quietly assumes one thing the harmonic drive refuses to give: a constant spring. Strain wave gears publish their stiffness as a three-zone curve — K1 from zero torque up to a threshold T1, K2 from T1 to T2, and K3 from T2 to rated torque — and the catalog numbers make the nonlinearity explicit.

For a size-25 cup-type component set (CSG-25-100, 100:1), Harmonic Drive's published curve reads K1 = 31,000 Nm/rad, K2 = 50,000 Nm/rad, K3 = 57,000 Nm/rad, with the zone boundaries at T1 = 14 Nm and T2 = 48 Nm against a rated torque of about 87 Nm (Harmonic Drive AG catalog data). The low-load zone boundary sits near 16% of rated torque — meaning the region where a robot spends much of its life, holding its own weight quietly, is precisely the region with the softest spring.

The arithmetic consequence for resonance is modest but real: with K1 about 1.8 times smaller than K3 (57,000/31,000 ≈ 1.84), the pole frequency at light load sits up to about 1.35× lower than at heavy load (resonance scales with √K). That is not a dramatic shift — but it moves the pole toward the controller bandwidth exactly when the joint is doing precise, low-torque work, and it means a notch filter tuned at one load condition can be mistuned at another. Tuttle and Seering measured this behavior early and put it plainly: static stiffness profiles on real units are "highly nonlinear" (Tuttle & Seering, "A nonlinear model of a harmonic drive gear transmission," IEEE Transactions on Robotics and Automation, 1996, doi:10.1109/70.499819).

And stiffness is only half of it. The same transmission contributes Coulomb friction and structural damping that vary with torque and speed, so the joint's dynamics are nonlinear end to end — flexibility, friction and damping coupled together, as Taghirad's thesis documents in detail. A joint tuned at one operating point is, strictly speaking, a different plant at another.

Source: Harmonic Drive AG published component-set data; single-vendor catalog figures, typical of the strain wave class rather than guaranteed for all units.

Why High Gear Ratios Pull Resonance Toward the Controller

Gear ratio enters the resonance problem twice, and both times it moves the pole in the direction controllers don't want.

First, load inertia is reflected to the motor side divided by N². That is the mirror image of the gear-ratio-squared tax on reflected inertia from the load's point of view — the same N² that shields the motor from the load also decouples the controller from the load side, and the joint's pole migrates toward √(K/JL) as N grows. When the flexible element is the gear itself (an elastic joint in Spong's sense), the pole formula ω = √(K·(1/JL + 1/(N²·Jm))) shows it directly: large N pins the frequency to the load-side term. A heavily geared humanoid knee, whose link inertia dwarfs everything behind it, owns a resonance set almost entirely by the gear's stiffness and the leg's inertia — and because the motor no longer feels the load, the encoder on the motor sees less and less of the ringing it must eventually damp.

Second, high ratio concentrates the entire compliance budget into one small component. Direct-drive joints spread their compliance over big, stiff structures and push their first mode well above the control band; that is part of why direct drive controls torque so transparently — and part of why it pays in motor mass. The geared joint buys torque density and buys, at the same time, a first mode sitting at a few tens of hertz, close enough to the achievable torque-loop bandwidth that every gain increase is a negotiation with the resonance.

This is the quiet reason "just raise the bandwidth" — the advice that ends many torque-bandwidth debates (see our earlier piece on what torque bandwidth honestly measures) — runs into a wall. Bandwidth is only useful up to a fraction of the first mode, and the first mode is set by components the motor engineer doesn't control and the datasheet doesn't list.

Torque Ripple: The Excitation You Already Paid For

A resonance is harmless until something excites it, and a geared joint ships with a built-in excitation source: torque ripple from the gear mesh itself.

In a strain wave gear, the teeth mesh in two zones simultaneously, so the dominant mesh vibration oscillates at twice the motor velocity — and its harmonics stack above that. Taghirad's measurements on a harmonic-drive bench make the spectrum concrete: at an input speed of about 156 rad/s, the torque spectrum shows distinct peaks at 312 and 624 rad/s, exactly 2× and 4× the input frequency. Tuttle and Seering's earlier experimental work found the same fingerprint on kinematic error, which appears "primarily at frequencies that were multiples of the wave-generator rotational velocity" (Tuttle & Seering, DETC1993-0319).

The danger is arithmetic. The ripple's base frequency scales with motor speed, so as the joint accelerates and decelerates through a working range, the ripple sweeps through a band of frequencies — and at some speed it will cross the first resonance. When it does, the transmission's own low stiffness does the coupling. Taghirad's summary table states the mechanism in one line: "torque ripples produced by gear meshing can interact with the low stiffness of the transmission to excite resonances," generating high vibration amplitudes in certain operating ranges. This is why a joint can be silent on the bench at one speed and audibly rough at another — the resonance hasn't changed, the excitation has moved onto it.

The same logic extends beyond the gear. Motor-side cogging and current-loop imperfections add ripple components of their own, and any of them that lands on the pole gets amplified. A joint whose resonance nobody has measured is a joint whose noisy operating points nobody can predict.

What Engineers Do About It

None of this is fatal, and none of it is new. The mitigation toolkit has four shelves, and mature joint designs use several at once.

Filter the loop. The notch filter is the industrial standard: identify the pole, place a notch on it, keep the gains. Drive vendors expose notch (and low-pass biquad) stages precisely for this — Galil's application note describes using notch filters "to control mechanical resonances that can destabilize motion control systems." The limitation is baked in: notches are fixed, and a resonance that moves with load (as strain-wave stiffness guarantees it will) can drift out of the notch. Multiple notches, adaptive tuning, or a conservative gain margin are the workarounds.

Shape the command. Instead of fighting the resonance in feedback, avoid exciting it in feedforward. Input shaping — convolving commands with a sequence of impulses timed to cancel the mode — originated with Singer and Seering's 1990 ASME paper "Preshaping Command Inputs to Reduce System Vibration" (J. Dyn. Sys., Meas., Control 112:76–82, doi:10.1115/1.2894142), refined shortly after by the vector-diagram approach of Singhose, Seering and Singer (ICRA 1990). For a robot, the technique lives naturally in trajectory generation: smooth the joint accelerations, and the ripple-and-resonance interaction simply never gets an invitation.

Feed back what the motor can see. Acceleration feedback adds active damping using the signal most correlated with the ringing. Zhang's two-inertia study pairs a notch with acceleration feedback and trims the modeled resonance to within 10 Hz of measurement; practical drives offer variants of the same idea. The catch returns with high-ratio gearing: the motor-side encoder is an imperfect window onto load-side oscillation, which is one more reason integrated joints increasingly instrument the output side.

Pick the spring deliberately. Stiffness is a design variable. A stiffer gear class (planetary and cycloidal families carry higher, flatter stiffness curves than strain wave units in comparable sizes) raises the pole away from the control band; a deliberately soft, instrumented spring — the series-elastic philosophy — converts the resonance problem into a measurable, damped one, trading control bandwidth for force fidelity. Pratt and Williamson's original SEA argument (IROS 1995, doi:10.1109/iros.1995.525827) is still the cleanest statement of that trade: lowering interface stiffness buys "more accurate and stable force control." And the strain wave's own friction, so often mourned as an efficiency loss, is quietly a passive damper that keeps the mode from ringing freely — Taghirad's whole program of "robust torque control" exists because flexibility, friction and damping must be modeled and compensated together, not because the joint is hopeless.

The Checklist for the Uncharted Spec

Resonance cannot be read off a datasheet, but it can be asked for. Before committing to a joint — or wondering why the last one buzzes — the questions that matter:

None of these questions is exotic. All of them are cheaper asked before the joint is in the leg than after.

The Mode You Know Is the Mode You Own

A humanoid lives in exactly the regime where resonance bites: high gear ratios, low-torque quiet phases, torque ripple swept across the band, and controllers asked for more bandwidth every generation. The joints that survive that regime are not the ones with the largest stall torque on the datasheet — they are the ones whose first mode has been measured, mapped across load, and designed around, with filters, shaping, sensing and gear choice all pointing the same way.

That is the engineering view we take to every module we build. The integrated joints in our humanoid robot actuator line and our lightweight joint modules are specified the way this piece argues: stiffness curves, not single numbers; dynamics first, torque second. And the two invisible specs that frame today's story are covered in depth in our earlier pieces on the gear-ratio-squared reflected-inertia tax and what torque bandwidth honestly measures.

The resonance will always be there. The difference between a joint that rings and a joint that sings is whether its builders put the mode on the map.

Source: Trigger discussion of harmonic-drive resonance as a hidden failure mode in Chinese industry commentary (唯能机器人润滑专家, WeChat, 2026-10-26); Taghirad, Robust Torque Control of Harmonic Drive Systems, McGill University, 1997; Tuttle & Seering, IEEE Trans. Robotics and Automation, 1996 (doi:10.1109/70.499819) and DETC1993-0319; Zhang et al., Mechanical Sciences 12:305–318, 2021 (doi:10.5194/ms-12-305-2021); Harmonic Drive AG component-set catalog data; Galil Application Note #3502; Singer & Seering, ASME J. Dyn. Sys., Meas., Control 112:76–82, 1990; Pratt & Williamson, IROS 1995. Catalog stiffness figures are single-vendor published values, cited as class-typical engineering data.