BLOGS

Torque density is the number every humanoid joint gets judged by, yet most comparisons quietly mix peak with continuous, and motor with module. A task-oriented co-design prototype from HKUST(GZ), published on arXiv in 2026, measured 35.7 N·m/kg at the output — roughly 60% higher than a widely used commercial geared joint actuator while being 18.6% lighter. The levers behind that number are not exotic; they are auditable engineering choices. This article walks through each one with the verified figures, then adds the lever that hides behind all of them: cooling.

First, Fix the Metric: Peak Versus Continuous

Before comparing any two joints, align on what the number means. Mechanical power is torque times angular velocity (P = T·ω), but robot joints spend their lives at low speed and high torque, so engineers compare torque density — torque per unit mass (N·m/kg) or per unit volume (N·m/L).

The trap is the rating basis. Peak torque lasts seconds; continuous torque defines what a joint can do all day. A module that quotes 120 N·m peak may sustain 40 N·m before its windings hit the thermal limit. Two joints with identical peak ratings can differ by a factor of two in useful, all-day output. Anyone who has followed our earlier walkthrough of duty-cycle-based joint sizing will recognize this as the same discipline applied to the datasheet itself: the continuous column, at the ambient temperature you will actually run, is the spec that matters.

Why Every Gram in a Joint Multiplies Upstream

A humanoid carries roughly twenty to thirty joint actuators, and they are stacked in series: the forearm joint lifts the payload, the upper-arm joint lifts the forearm plus the payload, and the hip and knee joints carry everything above them. Every gram added at the distal end is multiplied by every joint upstream of it — in inertia the whole robot must fight, and in battery drain the whole robot must carry.

This is why power density is not a tuning knob for humanoid joints; it is the first-order constraint. The recent co-design work from HKUST(GZ) made the point quantitatively: the team's task-oriented framework optimized the motor, gearbox and control jointly rather than component-by-component, and the fabricated two-DoF jumping-leg prototype reached a nominal rated output torque density of 35.7 N·m/kg — approximately 60% higher than a widely used commercial geared joint actuator, 18.6% lighter, and under matched bench conditions it jumped 12.0% higher at twice-rated torque (Source: arXiv 2609.22795, HKUST(GZ), 2026).

The 35.7 N·m/kg figure is a *rated output* torque density measured on the fabricated prototype — not a simulated value and not a peak rating. That is precisely what makes it a useful benchmark: it is comparable across modules in a way peak numbers are not.

Lever 1: Coaxial Integration and the Large-Gap-Radius Motor

The first lever is geometric. A complete joint module packs a motor, a reducer, a drive and encoders; if these sit in a scattered, stacked layout, the result is a long axial package, dead weight in couplings and housings, and transmission losses between every pair of components. Coaxial integration puts all of them on one axis, compressing axial length and cutting both losses and mass.

The physics behind the flat-and-large layout comes from the motor itself. MIT's biomimetics lab laid out the dimensional argument that made the Cheetah actuators famous: for a brushless motor of constant thickness, output torque scales with the square of the air-gap radius while mass scales linearly with radius — so torque per kilogram rises as you grow the gap radius (Source: MIT Biomimetic Robotics Lab, "Optimal Actuator Design"). Their implementation used a 77.5 mm gap-radius motor with a single-stage 5.8:1 planetary stage, and the resulting robot posted a cost of transport of 0.51 — still a landmark figure. Industry estimates put that Cheetah-era custom motor's pure motor torque density near 27 N·m/kg, which gives a sense of how far integration and co-design have pushed module-level numbers since.

The practical consequences for a joint module are direct: a large-diameter, flat motor lets the reducer ratio drop, and a lower ratio removes gear stages — each removed stage deletes its friction, its mass and its failure modes.

Lever 2: Reducer Miniaturization Without Backlash

The reducer is usually the heaviest and bulkiest element inside a joint, so it is where miniaturization pays the most. The engineering tension is that shrinking a gearbox usually worsens backlash — and backlash is what turns smooth commanded motion into positioning error and impact noise.

The way out is precision rather than compromise. Modern planetary design attacks backlash through tooth profile, gear grade and manufacturing accuracy rather than through preload springs that add mass and friction. Academic work on compact planetary stages has pushed measured backlash down to single-digit arc-seconds, and industry-standard precision grades (for example ISO 1328 grade 6 gearing) keep motion consistency while trimming surplus structure. The choice between strain-wave and planetary transmission for each joint position — and the hybrid maps real humanoids actually ship — we mapped in detail in our earlier reducer comparison.

When auditing a supplier's miniaturization claim, ask for backlash measured at the *output* flange, over the full load range, not the motor-side spec sheet value. The two numbers routinely differ by a multiple.

Lever 3: Topology-Optimized Structure

Structural housings exist to carry load — but under real stress conditions, large regions of a conventionally designed housing carry almost none of it. Topology optimization finds those regions and removes the material.

The documented gains are substantial. A peer-reviewed 2025 study on an industrial welding robot's upper arm used size, shape and topology optimization and found topology the clear winner: a 40% mass reduction on the upper arm, with total robot arm mass down 9.7% and all strength and stiffness requirements maintained (Source: IIETA, Mathematical Modelling of Engineering Problems, 2025).

For a joint module the effect compounds with the other levers. Less housing mass lowers the joint's own inertia, which speeds dynamic response on every acceleration and deceleration cycle. And there is a second-order effect most datasheets ignore, which the next section makes explicit: the housing is not just structure — it is the joint's radiator.

Lever 4: Slot Fill Factor — the Copper Budget

The first three levers remove mass and volume. The fourth adds working material. The metric is slot fill factor: the fraction of a stator slot's area occupied by actual copper conductor.

Industry reference ranges run about 0.35–0.45 for conventional round-wire windings and 0.60–0.80 for hairpin or formed windings. The gap is not cosmetic — in a fixed stator volume, copper fraction sets the current-carrying capability, and therefore the torque envelope. Moving a winding design from the round-wire range toward hairpin-class fill is one of the most direct routes to more torque from the same iron.

Two cautions keep this lever honest:

• Higher fill raises the ceiling, not the floor. Sustained output remains thermally limited; without the thermal design to carry the heat out, the extra copper only shows up in short peak events.
• Fill factor interacts with the electromagnetic design. The goal is torque at acceptable copper loss, not copper for its own sake — the motor constant (torque per square root of ohmic loss) is the quantity that actually predicts continuous performance.

The Hidden Lever: Cooling Decides the Continuous Rating

Here is the lever that never appears in the torque-density headline but decides the continuous column of the datasheet. Nearly all copper loss becomes heat, and continuous torque is set by how fast that heat leaves the motor — not by the electromagnetic capability of the design.

Sandia National Laboratories quantified the ladder precisely for robotic actuators. Their brushless-motor housing work measured, against a nominal design: +50% heat transfer from housing conduction improvements alone, +79% with forced air, and +107% with pumped water — and strategically spending energy on cooling returned net energy savings of 4–6%, because a cooler motor wastes fewer watts producing the same torque (Source: Sandia National Laboratories, "Improving robotic actuator torque density and efficiency through enhanced heat transfer").

This is also where lever 3 cuts both ways: the topology-optimized housing must shed heat even as it sheds mass. Integration makes the problem sharper still — packing motor, drive and reducer onto one axis raises heat density per unit of surface area. The joints that hold their continuous ratings in the field are the ones whose thermal paths were designed as deliberately as their magnetics.

From Levers to a System: What Buyers Should Audit

None of these levers works in isolation; the system-level result is what the 35.7 N·m/kg prototype demonstrates. The summary table below is the audit checklist we suggest when evaluating any integrated joint module — our own included:

The takeaway for humanoid and arm builders: when two datasheets quote the same torque density, the difference between them lives in these five rows. Ask how each lever was implemented, ask for the measurement basis, and the real ranking emerges quickly.

For the architecture behind how modern joints integrate these elements — sensing, compute and thermal design on one axis — see our engineering note on the integrated joint. To see how these principles land in a product family spanning harmonic and planetary transmission, browse the Humanoid Robot Actuator line and the Lightweight Joint Module series.

Source: arXiv 2609.22795 (HKUST(GZ), 2026); MIT Biomimetic Robotics Lab; Sandia National Laboratories; IIETA MMAp 2025; EYOU Robotics technical blog (Oct 2026).