What torque density actually measures

Torque density — also called specific torque — is output torque divided by the mass that produces it. It is the rotary cousin of the power-to-weight ratio that decides aircraft, and it matters for the same reason: a humanoid carries its own actuators everywhere it goes. A stationary industrial arm can bolt a heavy gearbox to the floor; a biped cannot. Firgelli’s engineering guide puts typical industrial servo actuators at 5–10 N·m/kg and argues that a humanoid joint only becomes viable above roughly 15 N·m/kg peak, competitive above 25, and state of the art above 35 — while linear actuators are judged on the equivalent specific force, with a target above 4,000 N/kg.

The catch is that “the mass that produces it” is negotiable, and vendors negotiate it aggressively. A bare frameless torque motor of the kind used in humanoid joints delivers only about 2–5 N·m/kg by itself, per TechBuzz China’s component tracker. The reducer multiplies that torque 6× to 120× depending on architecture, which is how a complete joint reaches tens of newton-metres per kilogram — but the reducer, encoder, driver board, housing and output bearing all add mass of their own. Whether a published figure divides by the motor’s mass or by the whole module’s mass changes the answer by several multiples.

The ladder: from industrial servo to state of the art

Very few humanoid makers publish both peak torque and module mass for the same part, which is why honest module-level anchors are rare and precious. The clearest one on the open market is CubeMars’ AK80-9, the quasi-direct-drive module descended from the MIT Mini Cheetah playbook: 9 N·m rated and 22 N·m peak from a 485 g package that includes motor, 9:1 planetary reducer, encoder and driver. Divide and you get roughly 45 N·m/kg peak and 19 N·m/kg rated — the same actuator, a factor of 2.4 apart, depending on which torque you quote. That gap is the thermal story: peak torque is a burst number, and the continuous figure is set by how fast the joint can shed heat.

The definition games vendors play

Unitree’s H1 page is the best worked example in the industry. It states 360 N·m of knee torque and a peak torque density of 189 N·m/kg for the M107 joint motor — a spectacular number, more than five times Firgelli’s state-of-the-art bar. Both figures can be true at once because they almost certainly divide by different masses: 189 N·m/kg is presented as a motor-level figure, and the page does not say whether the mass in the denominator includes the reducer, encoder, driver and housing that a working knee cannot do without. Neither figure is wrong; they are answers to different questions.

When you compare actuators — or robot spec sheets — four questions strip most of the marketing out of a torque-density claim.

• Peak or continuous? Peak is a seconds-long burst limited by magnetics and current; continuous is limited by heat. The AK80-9’s own numbers differ by a factor of 2.4. A robot that must hold a squat lives on the continuous figure.
• Motor or module? A bare frameless motor at 2–5 N·m/kg becomes a module at tens of N·m/kg only after the reducer — and the reducer, bearing, driver and housing can outweigh the motor. Module-level mass is the number the leg actually carries.
• At what ratio? Gearing up multiplies torque density but taxes back-drivability and reflected inertia. A 100:1 harmonic joint and a 9:1 quasi-direct-drive joint at the same N·m/kg are very different joints to control, as our QDD explainer lays out.
• With or without the sensor stack? Torque sensors, dual encoders and brakes are heavy. Vendors that quote the stripped unit look better on paper than the ones that quote the part you can actually bolt into a shoulder.

Why every gram compounds in a biped

Torque density matters more in a humanoid than in almost any other machine because the penalty is recursive. Add 500 g to a knee actuator and the hip must produce more torque to swing the heavier shin — so the hip actuator grows, the waist that carries the hip grows, and the battery that feeds them all grows. Aircraft designers call this the weight spiral; humanoid designers live it joint by joint, and it is why mass sitting far from the torso hurts most. It is also why tendon-driven hands relocate finger actuators into the forearm and why several makers pull knee and ankle actuators up the limb and reach the joint through linkages.

The stakes are counted in dozens of parts. A full-size humanoid carries roughly 25–35 powered joints, and actuators account for the majority of the robot’s material cost. That makes torque density a supply-chain number as much as an engineering one: the maker who hits the same joint torque with 20 percent less mass ships a lighter, longer-running robot at lower cost per unit — which is exactly the race the actuator supplier landscape is being run over.

FAQ

Sources