GaN Power Stages in Humanoid Robot Joints: What the Datasheets and One Measurement Say

Where the power goes in a humanoid

Texas Instruments' application brief on humanoids counts about 40 permanent-magnet servo motors in a typical body excluding the hands, with a dozen or more micro motors in each hand. Finger motors draw a few amps; hip and leg motors draw 100 A or more. The brief puts the driver on a ring-shaped board of 5 to 10 cm diameter inside the joint, which is the constraint the rest of this article is about: the inverter must fit in the joint, run from the robot's battery bus and shed its own heat there. [1]

The bus is usually near 48 V. Unitree's G1 runs from a 13-cell series pack of 9,000 mAh charged by a 54 V, 5 A charger; the page gives no watt-hour figure, and the 48 V class is inferred from the cell count. The H1 pack is stated as 15 Ah and 0.864 kWh with a 67.2 V maximum. Figure's 2024 release gives Figure 02 a 2.25 kWh pack; Apptronik lists hot-swappable Apollo packs of four hours each; Boston Dynamics' Atlas sheet gives four hours, a three-minute swap and a 1.5 hour charge with no capacity. The commercial joint actuators that these buses feed are rated at 24 or 48 V. [2] [3] [4] [5] [6]

A hip actuator at 100 A RMS peak on a 48 V bus is the sizing case for a leg driver; a wrist at 10 to 15 A is the sizing case for the small ring boards below. The MyActuator manual also gives the thermal limit that datasheets rarely state: the X8 holds 129 N·m stall at 43.8 A for 5 s with a 30 °C rise, which is the kind of figure a joint's duty cycle has to be planned around. [7] [8] [9]

What a GaN stage changes: frequency, dead time, capacitors

A silicon MOSFET joint drive typically switches at 8 to 32 kHz with a dead time around 1 µs between the two transistors of a phase leg; a GaN phase leg switches at 40 to 100 kHz with a dead time of 10 to 20 ns. Those two numbers are the whole argument. Shorter dead time means less distortion of the phase current near zero crossings; higher frequency means smaller current ripple, which lets the DC-link capacitor shrink from an electrolytic to ceramics and lets a low-inductance motor run without extra filtering. [1] [10]

The one peer-reviewed silicon-versus-GaN drive measurement found for this article is by Palma, Musumeci, Mandrile and Barba at Politecnico di Torino, IEEE WiPDA 2021. On a 36 V, 5 A RMS drive at 400 rpm they ran silicon at 20 kHz with 500 ns dead time and GaN at 100 kHz with 14 ns. Inverter efficiency was 98.5 percent for silicon and 98.2 percent for GaN. Motor efficiency rose from 66.3 to 73.1 percent and total drive efficiency from 65.3 to 71.8 percent, 6.5 points, with the sixth harmonic of torque reported as disappearing. The input filter went from a 2.7 µH inductor with a 660 µF electrolytic to no inductor and 44 µF of ceramics. [11]

That result is the correct way to read vendor efficiency claims: the GaN inverter itself was slightly less efficient at five times the frequency; the gain was in the motor, through lower ripple current and cleaner current near zero. TI's seminar material gives the capacitor side with numbers: 680 µF of electrolytic at 16 kHz against 170 µF of ceramic at 60 kHz and 110 µF at 100 kHz, with a current-ripple difference of 0.53 A RMS between 10 and 100 kHz at 10 A. Its white paper on audible noise attributes the sixth-harmonic whine of silicon drives to dead-time distortion at every 60 electrical degrees. [11] [10] [12] [13]

TI's 48 V reference design TIDA-010936 shows the loss side of raising frequency on GaN: about 9.66 W at 40 kHz and about 10.5 W at 80 kHz at 14.8 A RMS, close enough that the application brief calls the losses almost identical. The design guide's 99.3 percent peak efficiency at 40 kHz is a calculation, not a measurement, and the GaN device reached 117 °C at 15.6 A and 40 kHz with no heatsink against a 125 °C junction limit. A joint drive without airflow is therefore thermally bound well before it is electrically bound. [14] [1]

Reference designs sized for a joint

Two patterns in the table matter more than any single number. The small EPC boards are sized for wrists and fingers at 10 to 15 A; the TI designs reach the 16 to 60 A class that elbows and some hips need; and nothing in the table reaches the 100 A hip case without a separate high-current board. Second, every continuous-current rating depends on a heatsink assumption, which for a joint means the actuator casing, so the number that applies is the one measured with the casing, as EPC's 7 A against 15 A shows. [15] [16] [17] [18] [19] [14] [20] [21] [22]

The trade-off the vendor pages do not lead with

A GaN phase leg slews fast. Reviews of wide-bandgap drives report edge rates up to 50 V per nanosecond, and tests on a 7.5 kW motor up to 250 kHz and 40 kV per microsecond. At those rates the cable between inverter and winding behaves as a transmission line: reflected-wave overvoltage can reach twice the DC-link voltage at the motor terminals, and the partial-discharge inception voltage of the winding insulation falls by up to 38 percent when the rise time drops from 1,000 ns to 100 ns. The stress lands on winding insulation and bearings. [23] [24]

In a humanoid joint the cable is short, often inside the actuator, which limits the reflection; the insulation question does not go away, and a drive that switches at 100 kHz needs its EMI filtering designed for that frequency. Gate-drive complexity and device cost are also higher for GaN; no source found for this article quantifies either for a joint drive, so that is stated as a direction, not a number. [23] [1]

Power management: pack, bus, regeneration and heat

Commercial humanoid vendors publish runtime and swap times, not the architecture between pack and joint. Research platforms do. AGILOped puts two 26.1 V lithium-polymer packs in series for a 52.2 V bus feeding ten MyActuator X6-40 joints over CAN at about 4 kHz, with diode-isolated hot swap, a 12 V 200 W converter for electronics and a stated 1.5 to 2.5 h of runtime; it describes no battery-management system and handles thermal protection in software without numeric limits. The Berkeley Humanoid uses two DJI TB50 packs behind a power-manager board, EtherCAT at 1 to 4 kHz to the drivers and a 25 kHz driver loop, and does not model thermal derating. [25] [26]

Regeneration is the part of power management with a measured number. On the MIT Cheetah, Seok and colleagues reported that 63 percent of the negative mechanical work done by the motors was recovered to the batteries, with lithium-polymer charging efficiency above 95 percent, drivers drawing 60 A from a 100 V supply and a 20 kHz current loop. At 6 m/s the robot drew 973 W, of which 739 W, 76 percent, was Joule heating in the windings. That heating share, not the switching loss, is why the motor-efficiency gain in the Torino measurement is the number to watch: a cleaner current waveform lowers copper loss for the same torque. [27] [11]

Can humanoid robots work for eight hours? Battery, heat and charging

What to ask a vendor before choosing a joint drive

• Continuous phase current with the actuator casing as the only heatsink, at the robot's ambient, not at 25 °C on a bench.
• Measured losses at the switching frequency you will use; a calculated efficiency is not a measurement.
• Dead time and edge rate in volts per nanosecond, and the winding insulation rating of the motor it will drive.
• DC-link capacitance and type, and the bus-voltage ripple at peak current.
• Regenerative current path and the pack's charge-acceptance limit when the leg brakes.