BLOGS

A humanoid spec sheet will tell you how many actuators it carries, sometimes which reducers it uses — almost never which motors. That omission hides the most consequential split in humanoid actuation: the body and the hands run on two fundamentally different electric machines. Knowing which motor sits where explains a robot's mass, its heat and its control feel better than any peak-torque number on the brochure.

One robot, two electromagnetic philosophies

A humanoid carries 25 to 35 powered joints, and every one of them is the same recipe: an electric motor, a transmission that trades speed for torque, and sensing to close the loop. The transmission war — harmonic versus planetary versus roller screw — is fought loudly and in public. The motor war is not. Complete-robot spec sheets almost never name the motor inside, yet motor selection decides the joint's mass, its heat behavior, its controllability and a meaningful share of its cost.

Industry chain teardowns give the cleanest public picture of how the split works out in practice. A leading humanoid platform's second-generation robot carries roughly forty motors: 28 frameless torque motors spread across its 14 rotary and 14 linear body actuators, plus six coreless motors in each hand for finger-level control. One robot, two electromagnetic philosophies — and the division of labor is not arbitrary. Body joints move mass: hips, knees, shoulders, everything that lifts and carries. Fingers move almost nothing but themselves, dozens of times a second, along paths that must be silky rather than strong. Different jobs, different machines.

That is the design logic we apply across Ti5ROBOT's own humanoid robot actuators: pick the electromagnetic machine for the joint's actual duty, then build the module around it.

What "frameless" actually buys you

A frameless torque motor is a brushless permanent-magnet synchronous motor sold as two loose rings: a stator ring of copper windings and a rotor ring of magnets. No shaft, no bearings, no housing, no end caps, no encoder. The robot maker supplies all of that — the joint's own structure becomes the motor's frame, the joint's bearing becomes the motor's bearing, and the reducer bolts straight onto the rotor ring.

The concept is older than you might expect: the first frameless units were built in 1949 for missile guidance gimbals. What is new is a customer that wants nearly thirty of them per robot at automotive volumes. Supplier literature keeps returning to three reasons for the trouble:

• Packaging. Deleting the housing and shaft removes dead mass, and the motor can share the joint's own bearings. Fewer parts, shorter axial stack, less weight multiplied up the kinematic chain.
• The hollow bore. A large through-bore lets cables and encoder lines pass through the middle of the motor — which is how a humanoid arm routes its wiring through its joints instead of around them. In a robot with thirty joints, that routing channel is the difference between an arm that survives a year of cycling and one that chafes its harness to failure.
• Torque at low speed. Frameless torque motors are wound with high pole counts to deliver large torque at walking-pace joint speeds, with low torque ripple — which matters enormously when the transmission is a quasi-direct-drive stage of only 6:1 to 10:1 and almost nothing separates the windings from the limb.

One detail is easy to miss: voltage. Industrial frameless motors were historically wound for 300–600 V servo drives, while a humanoid runs on a battery bus at 48 V or below. Humanoid-grade windings are rewound fat and short to make torque at low voltage — one of the quiet reasons robot makers cannot simply lift a motor from the factory-automation catalogue and drop it into a knee.

The motor is not the actuator

Here is the number that surprises people. A humanoid-grade frameless motor, bare, delivers roughly 2–5 Nm/kg of torque density by industry-survey figures. A finished joint actuator module typically advertises 80–160 Nm/kg. Both numbers are honest. The gap between them is the reducer, which multiplies the motor's torque thirty- to a hundred-fold while adding less mass than the copper and magnets it lets you delete.

A single shipped product makes the point concretely: one commercially available 48:1 hollow-shaft planetary joint module weighs 1.4 kg and peaks at 222 Nm — about 159 Nm/kg peak for the assembly. The bare motor inside it could never approach that figure. The reducer does the multiplying.

The most common torque-density mistake in joint selection is comparing a bare motor's Nm/kg against a geared module's. Always ask two questions before believing any number: motor-only or complete module — and peak or continuous?

It follows that when a robot maker chooses a motor, the real question is what the transmission will forgive. A high-ratio harmonic joint can use a more modest motor and hide its ripple; a quasi-direct-drive joint with a 6:1 or 9:1 stage needs a fat, high-pole-count torque motor, because there is almost nothing between the windings and the world. That trade is the same one we mapped when comparing direct-drive and reducer routes for humanoid joints — the motor and the reducer are one design problem, not two purchases.

Bare-motor torque also scales steeply with diameter. Torque grows roughly with the square of the air-gap moment arm — motor engineers sometimes call it the D²L rule, where diameter counts for more than length — so a wider ring beats a longer one. That single piece of physics explains why humanoid joints keep trending toward large-bore, flat "pancake" packages rather than long skinny cans.

Radial flux now, axial flux next

Nearly every frameless motor shipping in a humanoid today is radial flux: the magnetic field crosses the air gap sideways, from a magnet ring to a winding ring. It holds the production benchmark for a boring reason — a century of manufacturing maturity.

The challenger is axial flux. Turn the geometry ninety degrees — a flat disc rotor facing a flat disc stator, the field running along the axis — and the motor gets 20 to 40 percent shorter for the same torque. Patent-landscape analyses put axial-flux torque density at 15–25 Nm/kg against roughly 10 for comparable radial machines, and argue that a disc motor mates naturally with pancake harmonic or cycloidal reducers for total actuator weight cuts of 40–50 percent. Treat those figures as the optimist's case: they come from vendors and patent filings, not teardowns of shipping robots. For now, the installed base remains overwhelmingly radial.

The pattern worth watching in both camps: torque-density gains are coming less from better magnets and more from geometry — wider bores, flatter discs, and motors designed together with their reducers rather than bought off a shelf.

The hands run on a different motor

None of the body-joint machinery fits in a finger. Hand actuation belongs to the coreless motor — the "hollow-cup" motor, in the Chinese industry's phrase. Its winding is not wound on an iron core at all: it is a self-supporting copper cup rotating around a stationary core, with no iron teeth for the magnets to grab. The payoff is zero cogging, almost no rotor inertia and smooth torque at very small scales — exactly what tendon-driven fingers need. In a typical dexterous hand, each coreless unit drives a tendon or micro-linkage through a small planetary stage, trading its high speed for finger-level grip force.

The engineering literature flags two honest weaknesses. First, a coreless motor can fail to start under a winding fault — with no cogging torque to nudge the rotor, a damaged winding can leave it dead in detent. Second, quality separation between the cheapest and finest units shows up in operating lifetime, not on the datasheet.

The economics are being rebuilt in real time. Swiss and German precision houses created this category for surgical tools and aerospace, and commanded premium pricing for decades; Chinese entrants now sell humanoid-grade units at 30–50 percent below Western benchmarks in the frameless class, and have pushed hand-motor costs down by an order of magnitude — the same structural cost collapse that already swept through reducers and roller screws. We tracked that dynamic from the module side in our lightweight joint module line: as component economics shift, integration quality — not component price — becomes the differentiator.

A buyer's checklist for the motor nobody names

Because spec sheets stay silent, buyers and integrators should ask four questions directly:

1. Which motor class is in the knee, and was it co-designed with the reducer? Co-designed motor-reducer pairs predict thermal endurance far better than the peak-torque figure on the brochure — a module that wins the peak number can still lose the shift. The difference between peak and continuous ratings is where joints live or die, as we laid out in the thermal case for integrated joint design.
2. Is the bore genuinely hollow? If cables route around the joint instead of through it, the "integrated" module will cost you a harness redesign.
3. Is feedback dual? Input-side and output-side encoders are what turn a strong motor into an accurate joint — the accuracy ceiling of a module is set at the output, as we explained in our piece on joint compute architecture and dual encoders.
4. Are the windings 48 V-native? A motor adapted from industrial servo stock rarely delivers the same continuous torque on a battery bus as one designed for it.

Ti5ROBOT's integrated joint modules follow exactly this recipe — a frameless-class torque motor matched to a precision reducer, dual feedback, hollow routing, and thermal margins sized for continuous duty — packaged as a single bolt-on unit.

Two motors, two jobs

The takeaway is simple. When you evaluate the next humanoid — or the next joint module for your own platform — ask what motor sits in the knee and whether it was designed together with its reducer. That answer predicts what the robot can do all day better than any number on the datasheet.

Source: humanoid.guide, "Frameless Torque Motors and Coreless Motors: The Electromagnetic Muscle Inside Humanoid Joints" (2026-10-12); CubeMars AKH70-48 product data; public industry-chain teardown analyses, 2026. Third-party platform and product figures are cited as one-line industry reference points.