The drive is the joint’s other half

A humanoid joint’s motor is a permanent-magnet synchronous machine (PMSM), and a PMSM produces no useful torque unless something feeds its three phase windings with precisely timed, precisely shaped currents. That something is the servo drive: a board, often no bigger than a matchbox, that sits inside or immediately behind the actuator. It reads the rotor’s position from the encoder, measures the current already flowing in each phase, runs field-oriented control (FOC) to compute the voltage vector that will produce exactly the demanded torque, and switches a bridge of power transistors to apply it.

Field-oriented control is the reason a modern joint can hold a cup of coffee without trembling. Instead of energising windings in crude steps, FOC mathematically rotates the measured phase currents into the rotor’s own frame of reference, splits them into a torque-producing component and a flux component, regulates each with its own control loop, and transforms the result back into three phase voltages. Torque becomes a number the controller commands directly — which is what makes current-based torque estimation possible in quasi-direct-drive joints, and what the torque sensors in high-ratio joints are there to verify.

Everything on that board has to fit where the mechanics leave room. Power Electronics News, surveying the field in 2025, notes that a wrist joint may leave an envelope of only 3–4 cm in diameter and 2 cm in thickness for its electronics, and finger joints less still. The drive also has to survive on the joint: it vibrates with every footstrike, bakes next to a motor winding, and must never push the robot’s outer skin past 55 °C — the temperature at which, as TI’s application brief bluntly puts it, full-thickness skin burns occur in 30 seconds of contact.

Why everybody lands near 48 volts

Bus voltage is a compromise between copper and chemistry. Go low and the currents become monstrous: TI calculates that a 4 kW leg drive running from a battery sagging to 39 V must carry roughly 102 A rms per phase — current that demands thick busbars, big transistors and serious cooling. Go high and every component needs more insulation and clearance, the battery needs more cells in series, and the robot crosses into voltage classes that trigger stricter safety rules for anything a human might touch. The industry’s answer, inherited from automotive 48-V systems and data-centre power, is a nominal 48 V bus that swings between roughly 39 V and 54 V across a battery’s charge cycle. TI expects 48 V to become the standard for humanoid power distribution, and recommends transistors rated at 60 V minimum — preferably 80–100 V — to absorb the switching spikes on top of a full battery.

Real robots match the pattern. Tesla’s Optimus carries a 2.3 kWh pack sized for a full working day, feeding 28 structural actuators. Unitree’s G1 runs on a 9,000 mAh quick-release pack good for about two hours of operation. And the loads those packs feed span three orders of magnitude: TI puts the drives in a humanoid anywhere from 10 W in a finger to 4 kW in a hip or knee, with most joints falling between 10 W and 1.5 kW. One battery, one voltage, forty to eighty motor drives hanging off it — plus a dozen more per hand.

The kilohertz ladder: loops inside loops

A joint drive is a stack of nested control loops, each running faster than the one above it. At the top, the robot’s whole-body controller sends torque or position targets over the network. The drive’s position and speed loops track those targets at 1 to 16 kHz. Underneath, the current loop — the one that actually does FOC — runs at 8 to 64 kHz, and the PWM that chops the bus voltage into phase voltages switches at 8 to 32 kHz on silicon designs. Each layer exists because the one above it can’t react fast enough: a footstrike or an unexpected contact arrives as a current disturbance long before it registers as a position error.

What actually lives on the board, then, is a fixed cast of components fighting for millimetres:

• A motor-control MCU or FPGA running the FOC maths, the loop stack and the safety monitors.
• A gate driver that translates the controller’s logic-level PWM into the fast, high-current pulses the power transistors need, inserting dead time so the bridge never shorts the bus.
• Six power switches — silicon MOSFETs or GaN FETs — arranged as a three-phase bridge.
• Current sensing in at least two phases, feeding the FOC loop at full rate.
• Encoder interfaces — magnetic, optical or resolver — for rotor angle and, in dual-encoder joints, output position.
• A network interface, CAN-FD or EtherCAT, tying the joint into the whole-body controller.

Silicon, GaN and the euro-coin drive

The power stage is where the newest material science enters the joint. Gallium-nitride transistors switch far faster and cleaner than silicon MOSFETs, and the numbers EPC quotes for humanoid joints are stark: a GaN inverter comfortably switches at 100 kHz where silicon designs run at 50 kHz, and its dead time — the safety gap in each switching cycle when no transistor conducts — drops to about 25 nanoseconds, a quarter of a percent of a 10-microsecond cycle, against the 2–5 percent a silicon bridge typically loses. Less dead time means less distortion in the phase currents, which means less torque ripple, less audible whine and less heat. EPC illustrates the space argument the same way: a conventional silicon driver stage occupying some 50 cm² of board shrinks, in GaN, to roughly the footprint of a euro coin.

The trade-off is cost and maturity, and it is narrowing fast. EPC’s EPC91104 reference drive for humanoid joints — a 14–80 V board delivering 14 A steady and 20 A pulsed from a single ePower stage IC — prices the GaN chip itself at about $2.50 in volume. TI’s Giovanni Campanella frames the silicon side of the ledger: simply doubling the switching speed of a silicon design costs 20–30 percent more in power losses, which in a battery-powered biped is runtime. None of this makes silicon obsolete — most shipping humanoids still switch silicon at 8–32 kHz perfectly well — but for the smallest, hottest, most torque-sensitive joints, wrists and hands especially, GaN is the direction of travel.

One bus, sixty nodes

A humanoid is a distributed computer whose peripherals happen to be joints. With 40–80 drives in the body and a dozen more in each hand, the wiring architecture matters as much as any single board. TI puts the minimum bandwidth for the whole-body control network at about 8 Mbit/s and names the two realistic candidates: CAN-FD, cheap and robust but limited in throughput, and Ethernet-based fieldbuses, EtherCAT above all, which daisy-chain through the limbs with microsecond-level cycle synchronisation. Networking vendors are already building for the latter: ASIX’s EtherCAT slave controllers target exactly this market, with multi-port junction chips that let a spine distribute one network into four limbs. Commercial joint modules increasingly hedge by speaking both — MyActuator’s RH-14 harmonic joint module, for instance, ships with EtherCAT and CAN interfaces on the same unit, and Schaeffler’s production-ready humanoid actuator integrates motor, gearbox, encoder and the controller in one 60–250 N·m housing.

FAQ

Sources