By Sarah Xiao 
A humanoid robot carries thirty to forty motors before anyone counts the hand. Tesla puts roughly 28 actuators in Optimus’s torso alone, plus a dozen or more micro-drives in the hands. They are not interchangeable parts with different labels — a shoulder has to hold up a hundred newton-metres, a finger has to settle a grip in milliseconds.
The industry converged on one answer anyway: frameless torque motors for the big joints, coreless motors for the fingers. It is worth being precise about why, because the usual explanation — “frameless is stronger, coreless is faster” — is a restatement, not a reason. Both motors win by removing something, and what they remove is exactly what the other application cannot afford to keep.
Two subtractions, not two products
A frameless motor arrives as a stator and a rotor. No housing, no bearings, no end caps, no shaft. The joint’s own structure supplies the bearing support and the sealing, and the rotor bolts straight onto the load.
A coreless motor removes the other thing: the iron core. The winding is wound into a self-supporting cup and spins in the air gap with nothing ferromagnetic inside it.
Same design philosophy, opposite consequences. Removing the housing gives you diameter you can spend on torque. Removing the iron gives you inertia you no longer have to accelerate.
The number nobody should compare
Here is where most spec sheets start lying, not by falsification but by omission of scope.
The Veichi FT1 series publishes 96.4 percent motor efficiency and a peak power density of up to 38 N·m/kg. Those are measurements of a bare motor under stated conditions. A high-performance integrated joint module, by contrast, starts above 100 N·m/kg of peak output torque density, and some in-house designs go considerably higher.
Those two figures have similar units and describe different objects. One is a motor’s power density; the other is a joint’s torque density measured after a gearbox. Placing them side by side invites the reader to conclude that a bare frameless motor can drive a load directly. It cannot. A frameless torque motor is a high-speed, low-torque device; everything a joint needs comes after the reducer.
That is why the reducer argument is a design argument, not a motor argument. A harmonic drive gives a high ratio and good accuracy, and its flexspline is the wear part. A planetary gearset takes shock loads better but asks more of the motor’s own torque density. Either way the frameless motor is the source of motion, and the reducer is part of the joint — never a substitute for the motor.
The hollow bore deserves its own line. Cables and encoder lines pass through the rotor centre, which is the difference between an axial stack that fits inside a knee and one that does not.
Diameter is the expensive axis
Torque in an electric machine scales with the square of rotor diameter and linearly with stack length. Double the diameter and you get four times the torque; double the length and you get two. That is the entire reason frameless motors exist: the joint housing sets the diameter, and nobody is willing to make the robot’s leg thicker.
The same scaling law has a second term that gets quoted far less often. Rotor inertia, for a solid rotor, scales with the fourth power of diameter. Double the diameter and torque rises fourfold while inertia rises sixteenfold.
| Double one dimension | Torque | Rotor inertia | Rotor mass |
|---|---|---|---|
| Rotor diameter ×2 | ×4 | ×16 | ×4 |
| Stack length ×2 | ×2 | ×2 | ×2 |
Scaling-law illustration for a solid cylindrical rotor, ignoring saturation and thermal limits.
Read the table right to left and the design constraint appears. Diameter is the only axis that buys torque faster than it buys mass — and it buys inertia far faster than either. For a hip that has to hold a pose, sixteen times the inertia is a control problem you can solve with a good controller. For a finger that has to stop on command, it is a control problem you cannot solve at all.
What the finger actually buys
Move the same reasoning down to the hand and the parameter order flips. A finger joint needs almost no torque, but it needs very low inertia, very fast response and very smooth torque.
A coreless rotor has no laminated iron, so its inertia is roughly a tenth of an iron-core motor of similar size. The immediate benefit is bandwidth. Here too the numbers need separating: industry material commonly cites mechanical time constants under 28 ms for coreless motors, with some products below 10 ms. The frequently repeated “under 5 ms” describes a flagship micro model under ideal conditions — it is not the class specification, and quoting it for a catalogue of coreless motors overstates what a normal finger joint will see.
The more durable advantage is cogging-free torque. An iron-core motor shows pronounced cogging at low speed, which is exactly the regime where a hand picks up an egg or turns a knob. Smooth torque at near-zero speed is not a luxury in those tasks; it is the task.
Lifetime is two different numbers
Coreless motors have a hard boundary, and it is where most comparisons go wrong twice.
The first error is treating a small motor as a strong one. Continuous output power is limited, temperature rises quickly, and single-unit torque is tiny — a coreless motor only becomes a finger actuator after a gearbox or a tendon drive.
The second error is quoting one lifetime figure for the whole category. Brushed coreless motors are limited by brush wear: Maxon’s data puts typical life at 1,000 to 3,000 hours. Brushless coreless motors commutate electronically and reach 10,000 to 20,000 hours, at which point the bearing becomes the limiting component. “1,000 to 3,000 hours” is the boundary of the brushed variant, not the fate of coreless motors in a dexterous hand.
And even the brushless figure may not be the binding constraint in production. In current dexterous hands, the parts that fail first are often the transmission elements — tendons and linkages wearing out within tens to hundreds of hours — while the motor is the component that fails last. Judging a hand’s life from its motor datasheet assumes the motor is the weakest link. Usually it is not.
Why you cannot swap them
The obvious follow-up questions both have short answers.
Can a frameless motor be shrunk into a finger? No. The iron in a miniature frameless rotor becomes inertia that the upstream joint has to carry, and dynamic bandwidth drops along the chain.
Can coreless motors be banked up to drive a hip? No. Torque and thermal budget both fail, and stacking units or adding a large reducer consumes the packaging benefit that made an integrated joint worth building.
The division of labour is not a preference. It is what remains feasible after four walls — diameter, inertia, heat and cost — are all applied at once.
The next variable is axial flux
The arrangement is not frozen. Axial-flux motors are further along than “long-term option” language suggests: products already sit in the 30–50 N·m/kg band for torque density, Tesla is reported to be exploring an axial-flux design around 30 N·m/kg for Optimus, and at the integrated joint level there are prototype figures above 200 N·m/kg peak torque density.
What axial flux offers is axial compactness at equal power, which matters enormously in a flat joint, plus light weight. What it costs is a tighter tolerance chain, real axial magnetic pull that the structure has to absorb, and manufacturing expense. The realistic near-term outcome is that axial flux takes positions where axial length is the binding constraint — flat, thin joints — rather than replacing frameless motors across every large joint.
On the finger side, the direction is clear even if the timing is not: brushless replacing brushed. The harder constraint before volume production is not the motor at all, but the durability of tendons and linkages under high-cycle work.
What to check on a spec sheet
| Number on the sheet | What it actually describes | What to ask instead |
|---|---|---|
| Peak power density, N·m/kg | Bare motor, stated conditions | Torque density at the joint output, after the reducer |
| Peak torque density, N·m/kg | Integrated module, often peak not continuous | Continuous torque and the duty cycle behind it |
| Mechanical time constant | Best-case sample, sometimes a flagship model | Distribution across the shipped range |
| Life, hours | Brushed or brushless, rarely labelled | Commutation type, and the life of the transmission |
| Efficiency, percent | Motor alone at one operating point | Efficiency across the joint’s real load profile |
Every row in that table is a question about scope. The motors are not the problem; the missing qualifiers are.
The unromantic conclusion
Neither motor wins a contest against the other, because they are not in the same contest. A frameless motor trades housing and bearings for torque per millimetre of diameter. A coreless motor trades an iron core for response per millisecond. Each is the optimum of its own physical accounting.
As long as humanoids need both a limb that carries weight and a hand that picks up an egg, that dividing line holds. And the discipline that makes the line legible is dull but decisive: know which object each number belongs to.
Sources: motor and joint figures from the original Chinese industry commentary and the manufacturers cited therein (Veichi FT1 series, Maxon lifetime data); axial-flux figures from published product ranges and reported Optimus design targets. The scaling-law table, the inertia argument and the spec-sheet checklist are Latest Robot News editorial.