BLOGS

Every humanoid joint datasheet quotes a peak torque. Almost none quotes a torque bandwidth — the frequency at which that torque can actually change. Yet bandwidth is what decides whether a robot softens a foot strike, keeps its grip when a box starts to slip, or yields gracefully when a person leans on its arm. Published figures for real legged-robot joints run from 8.5 Hz to 150 Hz, and the spread says less about hardware quality than about how each number was measured.

Torque bandwidth is a frequency, not a force. It is the point where a joint's delivered torque falls to about 71% of a sinusoidal command (−3 dB on a Bode plot) during a frequency sweep. Research joints have measured anywhere from 8.5 Hz to 150 Hz — but the number is meaningless without five pieces of test context: torque or position, open or closed loop, amplitude, load case, and sensor.

What Torque Bandwidth Actually Measures

The test itself is simple. Engineers feed the joint a sine-wave torque command and sweep its frequency upward — a chirp — while recording what the joint actually delivers. At low frequencies the output tracks the command; as frequency rises, the output shrinks and lags. The bandwidth is the frequency at which the output amplitude has dropped to about 71% of the command, the −3 dB point on a Bode plot.

Two things make the figure meaningful. First, what is being followed: torque, force, or position. Position bandwidth includes the inertia of the limb being swung, so it is almost always lower — the Blue arm from Berkeley reports 7.5 Hz for position but 13.8 Hz for torque on the same hardware. Second, what is being held still: many tests clamp the actuator output so only the actuator moves, not a free-swinging limb. Open-loop figures describe the hardware alone; closed-loop figures include the controller.

Why a Humanoid Needs More Than Its Walking Speed Suggests

A walking gait cycle lasts about a second — roughly two steps per second — and fast finger motions sit at 4–6 Hz, so it is tempting to conclude a joint needs little more bandwidth than that. The reasoning fails because bandwidth is not about tracking a planned trajectory — it is about reacting to what the robot did not plan for: a foot landing early, a load shifting in the hand, a person bumping into an arm.

A 2025 preprint benchmark proposal for humanoid actuation (not yet peer-reviewed) states the stakes directly: a joint with 5 Hz torque bandwidth cannot safely render compliant behaviours, and wrist and hand actuators need at least 5–10 Hz in torque mode. Its worked examples target roughly 8 Hz for an ankle in walking and 10 Hz for a wrist in manipulation.

Learned control sharpens the requirement further. A modern humanoid policy typically runs on the order of 50 Hz and sends joint position targets (sometimes with stiffness or impedance parameters), while each joint closes its own PD and current loops far faster. If the joint cannot realise the torque those targets imply, the learned policy is effectively commanding a slower, softer robot than the one it was trained on — and the mismatch shows up as degraded balance and contact behaviour.

The Chain That Sets the Limit

Every stage between the command and the limb adds lag, and the slowest stage wins.

• The drive's current loop comes first. In the MIT Cheetah leg it runs at 20 kHz and reaches 400 Hz of current bandwidth. The usual rule of thumb holds that a control loop must run 5–10 times faster than the bandwidth it is meant to deliver.
• The motor and reducer come next. Raising the gear ratio N multiplies torque, but the motor's reflected inertia grows with N² and friction grows with it. This is the mathematical root of why high-ratio joints are poor torque followers — the same N² tax that shapes the direct-drive versus reducer debate.
• An output spring is the biggest single lever. Wensing and colleagues modelled the Cheetah leg's rigid transmission at a natural frequency of 79.7 Hz; adding a series spring of 70 N·m/rad dropped the lowest natural frequency to 13.5 Hz — close to the 9 Hz closed-loop torque bandwidth published for the StarlETH quadruped's series elastic actuators.
• Heat is the final limit. At large command amplitudes the motor runs out of current or thermal margin before it runs out of frequency, so accessible bandwidth shrinks as torque amplitude grows.

The Published Numbers — and Why They Rarely Match

Researchers have published bandwidths for a handful of legged-robot and humanoid actuators, and they split cleanly along architecture lines: stiff, low-ratio quasi-direct-drive joints sit at the top, series elastic joints lower unless feedback pushes them up.

Source: compiled from Wensing et al. 2017; Paine et al. (Valkyrie); Kau et al. 2019; Yu et al. 2020; Isik et al. 2017; Gealy et al. 2019.

One detail deserves its own line: to our knowledge, as of late 2026, no major commercial humanoid maker publishes a measured joint torque bandwidth figure in its official specifications. The numbers above come from research groups with the methodology in the open.

Two Routes to a Fast Joint

The quasi-direct-drive route makes the hardware itself fast. With a ratio below about 10:1 and a stiff transmission, motor current becomes output torque almost instantly, and the joint can even estimate torque from current without a dedicated sensor. The cost is thermal: a low ratio means more motor current for the same output torque, so copper losses and heat management become the primary design constraint. This is the same physics that makes QDD joints run hot during static holds.

The series elastic route accepts a slow plant and fixes it with control. Valkyrie's elbow is passive at 13 Hz; with PD feedback on the measured spring torque it reached 70 Hz — 5.3 times its passive figure. But closed-loop bandwidth is paid for in current. When Valkyrie's knee tracking bandwidth was raised from 30 Hz to 82 Hz, power at the passive-bandwidth frequency rose 3.6-fold, from 27 W to 97 W. Higher gains at the same frequency mean more current, and that current becomes heat in the windings.

Bandwidth can also simply be traded away for cost. When UT Austin re-engineered its 18 Hz UT-SEA actuator for low-cost industrial use, the resulting SA-SEA reached 8.5 Hz against a 10 Hz target — a deliberate sacrifice documented in the paper.

The practical takeaway: a high bandwidth figure is never free. Ask what power the joint dissipates and what temperature its windings reach while delivering the quoted bandwidth. A number quoted at small amplitude with a grounded output and a cold motor tells you little about the same joint holding contact torque in a warm assembly cell.

Six Questions That Sort Out Any Bandwidth Claim

Because there is no standardised test, bandwidth figures are easy to flatter. Six questions separate an honest number from a marketing one:

The 2025 benchmark proposal suggests a common protocol: sweep from 0.5 to 60 Hz at about 10% of continuous-safe torque, and report gain and phase at 1, 5, 10 and 30 Hz. Until something like that becomes standard, the six questions above sort out most claims.

What This Means for Joint Module Buyers

For teams selecting integrated joint modules, torque bandwidth reframes the force-control conversation. A joint can carry a precision torque sensor and still be a poor force controller if its transmission inertia, friction, and thermal budget cap the bandwidth at a few hertz. The same logic applies in reverse: a low-ratio module with excellent bandwidth still needs the thermal headroom to hold contact torque continuously.

The practical checklist before committing to a force-control joint:

1. Ask for the Bode plot, not a single number. Phase lag at 5–10 Hz says more about stability in contact than the −3 dB point alone.
2. Ask for the load case — a grounded-output test removes the limb's inertia; a free-limb test is the harder and more honest one.
3. Ask for amplitude and temperature. Large torques and hot windings narrow the usable bandwidth.
4. Match the bandwidth target to the task: compliance rendering needs at least 5–10 Hz in torque mode; whole-body balance benefits from more at the ankle and hip.

This pairs directly with the error-budget view of joint accuracy — resolution and repeatability are static stories, while bandwidth is the dynamic half of the same conversation — and with the thermal case for continuous torque, because every extra hertz of closed-loop bandwidth is bought with watts in the windings. Force control moving into mass production makes the question urgent rather than academic: as joint modules ship with torque sensing as standard, the differentiator shifts from whether a joint can measure torque to how fast it can act on the measurement.

Source: humanoid.guide, "Torque Bandwidth in Humanoid Joints," October 2026; P. M. Wensing et al., "Proprioceptive Actuator Design in the MIT Cheetah," IEEE T-RO 2017; N. Paine et al., "Actuator Control for the NASA-JSC Valkyrie Humanoid Robot," Journal of Field Robotics; M. Sunbeam, "Human-Level Actuation for Humanoids," arXiv 2511.06796, 2025; K. Isik et al., Actuators 2017; D. V. Gealy et al., 2019; S. Yu et al., 2020; N. Kau et al., 2019.