Motors and joints

The joint is the robot. Motorised joints and hands are 50 to 70% of a humanoid's bill of materials estimate, and inside a rotary joint the gear reducer and the torque sensor each cost more than the motor. Heat in those joints is the real ceiling on what a cheap robot can do. If you understand one joint module, from magnet to output flange, you understand where the money, the margin and the failures in robotics live.

Inside a joint module

Real modules vary in which parts they keep. Unitree's pages list the choices it makes across G1, H2 and R1: low-inertia internal-rotor motors, industrial crossed roller bearings at the output, dual encoders per joint, hollow joints with internal wiring and local air cooling verified Sep 2026. A China Post Securities teardown of the G1 found a 4-in-1 small joint (motor, planetary reducer, encoder and driver) about 60 mm across, costing roughly CNY 300 for the reducer, CNY 200 for the motor, CNY 200 for the encoder, CNY 150 for structure and the rest for the driver board estimate. The G1 has no separate torque sensor in its small joints; it infers force from motor current, which works because its low gear ratio keeps friction low. You'll see why that matters below.

Motor physics in six numbers

Every robot joint motor today is brushless: permanent magnets on the rotor, copper coils on the stator, and electronics switching current through the coils in sequence so the magnetic field pulls the rotor round. Two names describe the same family. A BLDC motor is usually driven with blocky, trapezoidal current pulses; a PMSM is wound and driven for smooth sinusoidal current. In a robot joint, with a proper driver, the difference is mostly vocabulary. A servo motor is a motor sold with an encoder and a closed position loop; in industry that means a PMSM with a 17 to 25 bit absolute encoder and a full driver, US$500 to 3,000 per axis estimate.

The number that decides a robot's working day is heat. Current through copper wastes power as I²R, the copper loss. Double the torque and you double the current, which quadruples the heat. So a motor has two torque ratings. Rated torque (continuous torque) is what it can hold all day without cooking its windings. Peak torque is what it can give for a few seconds, typically 3 to 5 times continuous estimate, before the copper overheats. The allowed temperature rise of the windings sets the line between them. Spec sheets headline the peak. Your application lives on the continuous number.

Here's the arithmetic with illustrative numbers. A motor with K_t = 0.1 N·m/A at 20 A makes 2 N·m. Through a 10:1 reducer that's 20 N·m at the joint, less gear losses. To make 40 N·m you need 40 A and four times the heat. The honest figure of merit is K_m = K_t/√R, torque per square root of a watt of heat, because it says how much torque you get before thermals stop you.

Torque density, torque per kilogram of actuator (N·m/kg), is the number humanoid engineers chase, because every gram in a leg is a gram the motors above it must lift. Figure says its Figure 03 actuators run 2 times faster with improved torque density, without publishing the figure claim. Real joint torques from Unitree's pages: Go2 45 N·m, G1 knee 90 N·m (base) or 120 N·m (EDU), H2 leg joints 360 N·m and arm joints 120 N·m verified Sep 2026. EngineAI claims 450 N·m peak for its T800 unverified.

Magnets and where they come from

The rotor's strength comes from NdFeB magnets, sintered neodymium-iron-boron. Small doses of the heavy rare earths dysprosium or terbium keep them magnetised at joint temperatures above about 100 °C. A humanoid holds about 1 kg of NdFeB, with Apptronik's Apollo put at 2 to 3 kg contested, worth about US$50 to 200 per robot estimate. The cost is small. The control is total: China makes about 91 to 94% of the world's sintered NdFeB and about 99 to 100% of its heavy rare earths IEA; Adamas. China's April 2025 export licensing on seven heavy rare earths and magnets is still in force, and its October 2025 expansion is suspended until 10 Nov 2026 verified Sep 2026. Elon Musk said in April 2025 that the licence regime held up Optimus motor output corpus. Treat magnets as a supply risk rather than a cost line. Supply chain, money and geopolitics covers the controls.

Motor supply is scaling fast. A frameless torque motor costs CNY 2,000 to 5,000 (US$280 to 700) today, with a target below CNY 1,000 at scale estimate. Leadshine in Shenzhen delivered 120,000 frameless motors in 2025 and plans capacity for 2 million in 2026; Kinco shipped about 83,000 in 2025 estimate. UBS projects only about 30,000 humanoids shipped worldwide in 2026 estimate. Capacity is running far ahead of demand, and component price wars follow.

The driver and field-oriented control

At volume the driver is cheap. LeJu pays about US$46 per axis for driver electronics estimate, and Unitree puts the driver inside each joint. The power switches and control chips have historically come from TI, Infineon, ST, NXP, Renesas and onsemi; Chinese alternatives include Fortior Tech in Shenzhen for BLDC control chips and Geehy in Zhuhai for motor-control microcontrollers unverified.

Three loops inside every joint

A joint runs three feedback loops, one inside the other. The innermost is the current loop: FOC measuring phase currents and correcting them 10,000 to 20,000 times a second, which makes the motor produce exactly the torque asked for. Around it sits a velocity loop at 1 to 4 kHz, and around that a position loop at 100 Hz to 1 kHz typical rates. Each loop runs 5 to 10 times slower than the one inside it and trusts the inner loop to look instantaneous. Many humanoids skip the outer loops on the driver and send torque or impedance commands straight from the main computer at about 1 kHz. Control, from current loop to planner takes the stack further up.

Encoders: knowing where the shaft is

Bits turn into resolution by powers of two. A 14-bit encoder divides one turn into 16,384 steps, about 0.022 degrees each. A 17-bit encoder gives 131,072 steps, about 0.0027 degrees. Placement matters more than bits. A motor-side encoder sees fine, fast motion of the motor shaft but can't see backlash or twist in the gearbox. An output-side encoder sees the joint's true angle but coarsely. Premium joints carry both, and the difference between the two readings measures how much the gearbox has wound up, which gives you a torque sensor for free. On a G1 small joint the encoder costs about CNY 200, as much as the motor estimate. Heidenhain, Renishaw and Sick hold about 35 to 40% of the absolute encoder market between them estimate.

Feeling force: torque sensors versus current

Torque sensors are expensive and booming. A joint torque sensor is about 30% of a rotary actuator's cost estimate. Chinese makers shipped 70,000 joint torque sensor sets in the first half of 2025, against 50,000 in all of 2024 36Kr. China's humanoid makers bought about 12,300 six-axis force sensors in 2025, 92.8% from domestic suppliers, and one vendor, Landian Touch, held 72.6% of that market estimate. Wrist units cost CNY 15,000 to 38,000 and ankle units CNY 5,000 to 10,000, with a target near CNY 1,000 estimate. More in Sensors and perception.

Six ways to build a joint

Every humanoid maker has picked a side in one argument: how much gear reduction to put between the motor and the world. More reduction makes a small motor strong and precise, and makes the joint stiff and hard to push back. Less reduction needs a bigger motor, and gives a joint the world can push, which is safer and lets the motor feel contact. Backdrivability is the name for how easily the world can turn a joint from the output side.

That trade-off produces six families of joint. Most robots mix two or three.

The picks line up with each company's thesis. Unitree builds its own quasi-direct-drive joints because they're cheap, robust and let a robot feel the ground through its motors verified. Tesla split Optimus into 14 rotary actuators and 14 linear roller-screw actuators, putting the screws where shock loads land, in knees, hips and ankles AI Day 2022. 1X chose tendons and a 30 kg body so NEO is soft and quiet in a home, and builds its own motors on automated copper-coil winding lines in Hayward, California verified. Apptronik iterated through 35 electric actuator designs and claims 90%+ energy efficiency for Apollo 2's actuators claim. Figure builds its actuators in-house and publishes neither the rotary and linear split nor the reducer type not public. Boston Dynamics' electric Atlas has joints with continuous 360-degree rotation, and Hyundai Mobis will supply its actuators; Mobis's release says actuators are more than 60% of a humanoid's material cost verified. Agility's Digit 5 is reported to use cycloidal actuators in its legs for repetitive lifting unverified. Gears and screws covers the reducers and screws in depth.

What a joint costs and where the money goes

Motors get the attention and are the cheap part. In a rotary joint the harmonic reducer is about 36% of the cost, the torque sensor about 30% and the motor about 13.5% estimate. In a linear joint the roller screw is about 64% and the motor about 7%. Multiply by 28 joints and the actuators are about 56% of an Optimus bill of materials; the 14 linear actuators alone cost about US$28,000, around nine times the compute module Morgan Stanley. A high-torque leg actuator runs US$1,000 to 5,000 estimate.

Prices are falling fast. LeJu's joint module plus driver fell from US$433 to US$202 per axis in three years as its purchases grew 100 times, and its supplier targets US$140 for the mass market estimate. At retail the same class of module costs US$385 to 480 from DAMIAO or MyActuator verified Sep 2026. That spread between a US$202 volume price and a US$1,440 premium module is the whole Chinese cost advantage in one number. The cost chapter follows it through the whole BOM.

How joints fail

Four things kill joints: heat, encoder faults, bearing and gear wear, and water or dust. Heat comes first on cheap humanoids. The chain runs from software to copper: the controller's gains set the torque it demands, torque sets current, current sets I²R heat in the windings, and the windings have a hard temperature limit.

The public G1 reports show the pattern. A lab running reinforcement-learning policies on a 29-DOF G1 EDU saw violent motion in unfamiliar states, a broken left wrist pitch gear and right wrist "winding overheated" errors that shut down the whole arm owner report. Another owner fitted a heavier Inspire hand, raised the shoulder gains, and watched the shoulder pitch motor overheat until the arm stopped responding owner report. A burst of joint torque when switching control policies can trip the battery's over-current protection and drop the robot maintainer reply. Most humanoid hardware failures in public reports start in software.

The other three killers are slower. Encoders miss counts under electrical noise or overspeed, get contaminated, or pick up interference when their wires run unshielded next to motor phases; an incremental encoder that misses counts is silently wrong until the robot re-homes. Bearings and gears wear, and grease dries out. Fanuc's Zero Down Time service watches reducer wear on every axis through motor torque and flags an abnormal reducer several weeks before it fails verified, which tells you the maker already owns the best predictive-maintenance data. Water, oil, sand and food debris get in: Gausium's warranty excludes their ingress, and Pudu's KettyBot is rated only IP20 verified. How robots fail and get fixed has the field data.

Terms in this chapter