Why humanoid robots overheat: the thermal cliff

Ask a humanoid robot to hold a five kilogram box at arm’s length and do nothing else. No walking, no reaching, no dramatic backflip. Just stand there. Depending on the machine, it has somewhere between two minutes and twenty before a shoulder or an elbow starts logging temperatures that force the controller to back off. The robot has performed no work in the physics sense, since nothing moved. It has still spent a large fraction of its usable duty cycle.

This is the constraint that gets the least attention in a field currently arguing about batteries and about whether the robot can fold laundry. Runtime is usually framed as an energy-storage problem, and that framing is comfortable, because energy storage improves on a predictable curve and can be solved by swapping packs. Heat is not like that. It is a rate problem, it lives inside the joints rather than in a removable module, and it puts a hard ceiling on what a given actuator can be asked to do indefinitely, no matter how much charge sits in the chest.

The square in the middle

An electric motor makes torque roughly in proportion to the current flowing through its windings. Push twice the current, get twice the twist. That is the useful relationship, and it is linear, which is why it feels intuitive.

The waste heat is not linear. A winding is a long coil of copper with some resistance, and the power it dissipates as heat is the current squared multiplied by that resistance: P = I²R. Doubling the torque therefore quadruples the heating. Tripling it costs nine times. This single asymmetry is the reason a robot arm that lifts a heavy object briefly and impressively cannot hold that same object patiently.

Two numbers follow from it, and every actuator datasheet carries both. Peak torque is what the joint can produce for a short burst, limited only by how much current the electronics will deliver and by how hard you can push the magnets before demagnetising them. Continuous torque is what it can produce forever without the winding insulation cooking, and it is set entirely by how fast heat can escape into the surrounding metal and air. In a conventional air-cooled actuator the continuous figure typically lands at something like a quarter to a third of the peak. Engineers describe the gap as a thermal cliff.

The consequence is uncomfortable when written plainly. If a robot needs 100 newton metres at the knee to rise out of a squat, and its knee actuator sustains 30, then any posture that demands more than 30 is a posture the machine is renting rather than owning. It is drawing down a thermal budget that has to be repaid by cooling off, and the repayment happens far more slowly than the spending.

Gravity never lets go

Wheeled machines mostly avoid the problem. A rolling robot at rest draws almost nothing, because a wheel bearing holds the vehicle up and the motors can be de-energised. The load path goes through structure, not through torque.

A legged machine standing upright is a different arrangement. Its mass is held off the ground by a stack of joints, each of which is a motor being commanded to resist a moment. The knee is not locked, it is being actively pushed against. Take away the current and the machine collapses. So gravity, which does no work and never gets tired, is converted at every instant into copper heating in the legs, and then again in the shoulders and wrists whenever the arms are out in front of the body, which is where useful arms spend their time.

This shows up as an oddly specific failure in practice. Teams moving control policies from simulation to hardware report robots that walk fine and then cook their wrists, because a wrist actuator is small, buried, thermally isolated, and asked to hold an offset payload for minutes at a time. The joints that overheat first are rarely the ones doing the athletics.

There is a further multiplier hiding in the drivetrain. Humanoids get their torque from small fast motors behind reduction gearing, because a motor big enough to produce joint torque directly would be too heavy to carry. Gearing multiplies torque at the output but does nothing about the fact that the heat is generated in a small, densely packed motor upstream, in a housing that also contains the gearbox, its lubricant and often the drive electronics. The heat concentration in these compact joint modules is exactly what makes conventional cooling inadequate: there is no room for a heatsink and no airflow to speak of.

What a body does instead

The comparison with biology is not flattering, and it is instructive in a way that is not the usual one.

Muscle is not thermally superior. It converts roughly a quarter of its chemical input into mechanical work and dumps the rest as heat, which is not obviously better than a good electric motor. It also has a genuine advantage in cooling, since it is perfused with blood and the whole body is an evaporative radiator, a trick no robot currently performs.

But the real difference is architectural. A standing human is not holding a squat. The skeleton stacks so that the line of gravity passes close to the joint centres, ligaments and joint capsules take load at the end of range, and the soleus keeps the whole thing upright with a low, intermittent, postural contraction. Standing costs only a few percent more than lying down. Tendons store and return energy during walking so the muscle does less net work per step. The body is full of passive structures that hold position for free.

Robot joints have almost none of this. Some designs add mechanical brakes for parking and springs to offset static load, but a general-purpose humanoid that must be back-drivable and compliant for safety cannot lock itself rigid the moment it stops moving. The price of behaving safely around people is paying continuously to stand near them.

Where this stands in 2026

Three lines of work are running at once, and they attack different parts of the problem.

The first is to move the heat out physically. Liquid cooling of proprioceptive actuators is an old idea from research bipeds, where circulating fluid through the actuator housing was used to raise sustained torque density well beyond the air-cooled figure. The newer variant is evaporative. A 2026 study of phase change hydrogels packed around joint motors reports a material that holds a large water fraction, releases it as vapour when the motor heats, and still carries several megapascals of compressive stress, so it can sit inside the load-bearing structure rather than beside it. That is a robot sweating, with the same catch a sweating animal has: the water is consumed and has to be replaced.

The second is to teach the controller that temperature exists. Until recently a locomotion policy learned in simulation had no notion of a thermal state at all, and would happily choose gaits that were efficient in torque error and ruinous in heating. Two 2026 papers close that gap by putting a thermal model of the motors inside the training loop. One adds motor temperature to the policy’s observations and penalises approaching the limit. The other trains a corrective residual policy on top of a normal locomotion policy, and reports the clearest number in this whole area: carrying a three kilogram payload, the thermally naive baseline overheats in about five minutes, while the thermally aware version keeps walking for more than thirteen.

The third is scheduling. If a fleet cannot avoid the cliff, it can be operated around it, with duty cycles that alternate loaded and unloaded tasks and units rotated out before they saturate. This is the same logic as rotating robots out to charge, applied to a quantity that is harder to see. Work on thermal recovery in multi-limbed robots treats posture selection itself as a cooling strategy: choosing which limbs bear load, and for how long, so that no single motor stays at the top of its range.

What none of this fixes

Every remedy has a mass penalty, and mass is torque. Coolant, pumps, plumbing and thicker housings all add weight that the legs must then hold up, which raises the continuous torque demand that the cooling was installed to relieve. The net gain is real but smaller than the component-level numbers suggest, and it only shows up honestly in a whole-machine measurement that few published results include.

The temperatures being controlled are also mostly not measured. A sensor can be placed on a housing, but the part that fails is the winding insulation deep inside, and its temperature is inferred from a model of current, resistance and thermal capacitance. That model drifts as the machine ages, as lubricant thins, as dust fills a vent. A thermally aware policy is only as safe as its estimator, and a confident estimate of a wrong number is worse than no estimate at all.

Then there is the honest accounting of the trade. A policy that survives thirteen minutes instead of five is not doing the same thing more efficiently. It is walking differently, generally more conservatively, and the endurance is bought with performance the operator wanted. That may well be the right trade, but it is a trade, and the demonstrations that report the endurance gain do not always report what was given up.

The deeper point is that this is not a transitional annoyance waiting on better magnets. The square in P = I²R is not going anywhere, and neither is gravity. Machines that carry their own power and hold their own weight with active joints will always be paying to stand still. The interesting question is not how to remove that cost but where to put it: in exotic cooling, in smarter control, in mechanical latches and springs that borrow from skeletons, or in accepting that a general-purpose body is a poor way to hold a box in one place, and that the robot should put it down.

Further reading

• Learning to balance motor thermal safety and quadrupedal locomotion performance, the residual-policy result and the five-versus-thirteen-minute comparison
• Learning thermal-aware locomotion policies for an electrically-actuated quadruped, the same problem attacked through the reward function
• Control of a high performance bipedal robot using viscoelastic liquid cooled actuators, on what active cooling buys at the joint
• Thermal recovery of multi-limbed robots with electric actuators, on treating posture and scheduling as cooling strategies