Peak torque is a sprint. Continuous torque is a shift.

Peak torque answers a narrow question: can the actuator produce one short, high-output event? It is limited by magnetics and drive current and is sustainable for seconds. Continuous (or rated) torque is limited by heat: it is the highest output the motor can hold indefinitely without its winding temperature running away. In air-cooled humanoid actuators the continuous figure typically lands at just 25 to 30 percent of the peak figure — a gap engineers have started calling the thermal cliff, because it is where marketing meets physics.

The published ratings of the quasi-direct-drive actuators used across research humanoids and Unitree-class robots show the gap plainly. CubeMars’ AK80-64, a common knee choice, is rated 120 N·m peak against 48 N·m continuous. The AK60-39 shoulder unit manages 72 N·m peak but only 24 N·m rated — a 3:1 ratio. For sizing a real joint, neither number is the one that matters most: what matters is the RMS torque of the actual motion cycle, computed over a full task and compared against the rated value with margin for hot ambient air inside the limb.

Where the heat comes from — and why it can’t get out

Motor heat is mostly resistive: power lost as heat scales with the square of current (P = I²R). Double the load on a joint and the heat roughly quadruples. Worse, copper’s resistance rises about 0.4 percent per degree Celsius, so a hot winding wastes more power than a cold one carrying the same current — a positive feedback loop that ends in thermal runaway if nothing intervenes.

The cruel part for humanoids is that the worst case is not running — it’s holding still. A motor holding a static pose produces torque at zero speed, so all electrical input becomes heat and none becomes mechanical work. A deep squat, a held box, a braced drilling posture: these are stall conditions, the most punishing state an electric motor knows. And unlike an industrial arm with a fan and open air around it, a humanoid actuator is sealed inside a limb — a totally enclosed, non-ventilated package whose heat must conduct out through every interface in turn: copper windings, stator, housing, limb structure, and finally the ambient air.

The numbers robots publish — and the one they don’t

Spec sheets reflect the incentive. Unitree publishes peak joint torques for the H1 — about 360 N·m at the knee, 220 N·m at the hip, 59 N·m at the ankle — and a peak torque density of 189 N·m/kg, along with a note that its low-inertia internal-rotor PMSM design improves heat dissipation. The G1’s knee is quoted at 90 N·m on the base model and 120 N·m on the EDU version, with local air cooling to keep motors in range. What almost no humanoid maker publishes is the continuous rating of the mounted joint at a stated ambient temperature — the number that predicts whether the robot can do a shift, not a somersault.

Firgelli’s engineering guide works the consequence through a concrete case: a knee actuator holding a 20 kg load in a bent pose needs about 80 N·m, but its derated continuous capacity inside a 40°C sealed leg might be closer to 18 N·m. Meeting the pose means running at more than four times rated current, which multiplies heat generation roughly nineteen-fold — and puts the winding at its thermal limit in under two minutes. That arithmetic is why demo videos are short, and why architecture choices differ: Agility’s Digit uses series-elastic actuators that store and return energy each step, trimming the continuous thermal load, while quasi-direct-drive designs accept high current draw in static poses in exchange for transparency and impact tolerance.

Duty cycle compounds it. At a modest 1.4 steps per second, a walking humanoid puts roughly 40,000 load cycles through its leg actuators in an 8-hour shift, each step carrying two to three times body weight. Heat that accumulates a little faster than it escapes on each cycle shows up an hour into the shift, not in a 30-second demo — which is why serious buyers now ask for thermal soak test data, not peak torque.

The cooling ladder: buying back continuous torque

Since the peak number is set by magnetics, the practical way to raise a joint’s continuous rating is to remove heat faster. Each rung of the cooling ladder buys a bigger usable fraction of peak: natural convection through a sealed housing sustains perhaps 15 to 20 percent of peak torque; forced air, where a fan and airflow path exist, 25 to 35 percent; liquid coolant channels in the motor housing 50 to 60 percent; and direct oil immersion of the windings 60 to 70 percent. Moving from passive air to liquid can therefore triple the work a joint sustains — without changing the motor.

The industry is climbing the ladder. Industry analyses of Tesla’s Optimus describe liquid cooling loops with 2–3 mm channels serving joint motors whose peak electrical loads exceed 500 W, holding core temperatures far below what a passive design would allow — figures Tesla itself has not confirmed, but consistent with the thermal math of an eight-hour work target. The idea is older than the current wave: researchers at the University of Tokyo’s JSK lab demonstrated evaporative water cooling of humanoid actuators a decade before Optimus existed, roughly doubling usable motor output, and the research frontier has now reached phase-change hydrogel cooling for joint motors, published in 2026. Air still rules the mid-market — fan-cooled compartments and airflow paths designed in from the start, as thermal component makers like NMB advocate — but every roadmap that ends in “full shift, full payload” runs through liquid.

How to read a spec sheet like a thermal engineer

For anyone evaluating a robot or an actuator, the thermal questions are more revealing than the torque table. The ones that matter:

• Ask for the continuous rating of the mounted joint, at a stated ambient temperature — not the motor’s bench rating in free air. The same actuator derates hard inside a sealed 40°C limb.
• Size against RMS torque over the real task cycle, not against peak. A motion profile’s RMS torque must sit below the rated value with margin; peak events only need to clear the peak rating briefly.
• Probe the static cases. Holding poses are stall conditions. A robot that lifts a 20 kg box is answering a different question than a robot that holds one while a person fastens a strap.
• Ask what happens at the limit. Well-engineered joints throttle or shed load on winding-temperature feedback before insulation damage (~130–150°C) or magnet demagnetisation; the spec sheet should say where sensors sit and what the derating curve looks like.
• Ask for soak data, not sprint data. An hour of the target duty cycle at target payload, with housing temperatures logged, reveals the slow heat accumulation that a demo never will.

FAQ

Sources