The case for GaN in humanoid actuators is usually made wrong. It is presented as an efficiency story, and efficiency is the weakest version of it. The real argument is that GaN makes switching frequency nearly free, and everything a humanoid joint actually needs follows from spending that frequency: finer current resolution, shorter dead time, smaller bus capacitance, and a power stage that fits inside the joint.
The loss curve matters more than the loss number
Start with the result that decides the architecture. Texas Instruments tested a 48 V GaN inverter reference design, TIDA-010936, at 40 kHz, 60 kHz and 80 kHz PWM and reports that the board losses are almost identical at 40 kHz and 80 kHz [1]. On TI's own measurement, doubling the switching frequency barely moves the loss.
Compare that to silicon. TI puts the penalty for a MOSFET based servo driver at 20 to 30 percent higher overall loss for a move from just 10 kHz to 20 kHz, and calls that unacceptable for humanoid robots [1]. That is one doubling of frequency costing between a fifth and almost a third of the thermal budget, in a joint that has nowhere to put the heat.
The two device families are not competing on the same axis. Silicon offers a fixed frequency and asks you to live with it. GaN offers a range and charges almost nothing to move inside it. Under a hard thermal ceiling, the second property is worth far more than a point or two of peak efficiency.
What you buy with the headroom
Current resolution. The current loop in a humanoid joint runs at around 8 kHz to 32 kHz, and the switching frequency is typically the same as the current loop [1]. TI's position sensor guidance quotes the same 8 kHz to 32 kHz band for humanoid joint sample rates [2]. That is a control period of 125 µs down to 31.25 µs. TI's own comparison shows motor line current at 10 kHz PWM carrying visible ripple, while the same waveform at 100 kHz resolves into a clean sinusoid with smaller ripple [1]. Torque quality in a robot that has to place a hand gently is a current waveform problem before it is a control law problem.
Dead time. The underrated one. TI puts GaN dead time in the 10 ns to 20 ns range, against about 1 µs for the silicon MOSFETs in these stages [1]. We ran the arithmetic: against a 31.25 µs period with complementary PWM, a microsecond on each of two edges is 6.4 percent of the period, and 20 ns is 0.13 percent. Duty cycle error of that size becomes current distortion and then torque ripple, and control tuning never fully recovers it.
The mechanism underneath is that GaN HEMTs have no body diode. Freewheeling happens through third quadrant channel conduction, so there is no reverse recovery charge and no diode recovery ringing on the switch node [1]. Our read of that: third quadrant conduction costs more voltage drop than a silicon body diode, so short dead time is not a bonus with GaN, it is a requirement.
Bus capacitance. The most consequential system result gets one sentence. Raising PWM frequency from 20 kHz to 80 kHz let TI replace the electrolytic bus capacitor with a ceramic of equal capacitance at a similar bus voltage ripple, and TI credits the ceramic with smaller size, longer life and better high frequency behaviour [1]. On a ring PCB of 5 cm to 10 cm diameter [1], the electrolytic is usually the tallest component. Removing it changes the height budget and the service life together.
The area number, read carefully
TI compares a MOSFET reference design at 720 mm² of power device area against a GaN design at 320 mm², a reduction of more than 50 percent [1]. Worth understanding what is inside that comparison. TI attributes it to two things: the lower specific on resistance of GaN, and the gate driver integrated with the half bridge in one package [1]. The LMG2100R026 packs a half bridge plus its driver into a device rated for 55 A continuous, and the 4.4 mΩ variant fits into 4.5 mm by 5.5 mm [1].
Both contributions are real and both are available to you. But they are separable, and a design team evaluating discrete GaN plus an external driver should not expect the full number.
Scale is the reason any of this matters
A humanoid carries about 40 servo motors through the neck, torso, arms, legs and toes, before you count the hands, where a single hand can hold more than a dozen micro motors [1]. Current demand spans from a few amps in a finger to 100 A or more in a hip or leg [1]. Robotics power levels as a whole run from 10 W to 4 kW and above on a common 48 V rail [3].
Multiply any per axis thermal saving by 40 axes and it stops being a component decision. Multiply an electrolytic capacitor's failure rate by 40 axes and it becomes a dominant reliability term in the robot. This is why joint level power electronics converges on GaN faster than adjacent markets do: the axis count amplifies every per unit advantage and every per unit failure mode.
What we would actually argue about
The open question is not whether GaN wins at 48 V. It is where to stop. TI describes the switching frequency as tracking the current loop at 8 kHz to 32 kHz, tests the reference design at 40 kHz to 80 kHz, and illustrates the waveform benefit at 100 kHz [1]. Those reconcile only if you decouple PWM frequency from the control update rate and run several PWM periods per control tick, which changes how you sample current and how you configure sigma delta decimation. The exchange rate between decimation ratio, resolution and latency is quantified in TI's 48 V current sense work [4].
That decoupling is the design move worth having an opinion about. It is also the point where the current sensing chain, not the transistor, becomes the limit. On our own 60 V board that is exactly where the tuning time went.
References
- Texas Instruments, Application of GaN FET in Humanoid Robots, SNOAAB3, January 2025.
- Texas Instruments, Choosing a Position Sensor in Motor Control, SLYT867, Analog Design Journal 2Q 2025.
- Texas Instruments, Cobots to Humanoids: Driving System Efficiency and Safety into Higher-Power Robots, SSZTD67, November 2024.
- Texas Instruments, High Resolution, Small Form Factor Phase Current Sense for 48V Robotics and Servo Drives, SBAA666, February 2025.