Humanoid robots are gearbox-limited
A humanoid robot demo needs to survive three minutes on a stage. A humanoid robot product needs to survive a shift. The gap between the two is usually described as an AI problem. It mostly isn't. Write down what the machine itself has to do and you get a short, brutal spec sheet:
Last a working day on a battery it can carry. A humanoid is an energy system with legs: every watt a joint wastes is battery mass it must haul around, which costs more watts. Cost of transport — the energy burned moving each kilogram a meter — decides whether the robot works a shift or spends half of it charging.
Carry something. Payload is the spec that makes a robot useful, and it is decided joint by joint: the torque each elbow, shoulder, knee, and hip can add beyond moving the robot's own body.
Survive. A step every half second is millions of load cycles a year through every leg joint — each footfall an impact, not a smooth ramp.
Be safe next to people. A machine built for human spaces must yield when it touches someone — as a property of the hardware, not only of software.
Be affordable, thirty times over. A full-size humanoid carries thirty-plus actuated joints. Joint economics are robot economics.
Walk each problem down the stack
Not in the motor: motors are mature, and the standard scaling argument (Ben Katz's motor scaling laws) shows that at fixed output torque and power, motor size barely depends on gear ratio across the practical range — the ratio relocates problems rather than removing them. Not in the battery: humanoids ride the EV industry's curve. Not in compute, which is sprinting. Trace each challenge to its bottleneck:
- Battery life → transmission losses, in both directions. A lossy drivetrain wastes charge twice over: directly, in its own friction, and indirectly, by dragging the motor away from its efficient operating region — extra current for the same output. And walking is full of negative work (the knee absorbs at every heel strike): a drivetrain that passes energy backward cleanly recovers it; one that doesn't throws it away.
- Thermal headroom → the same losses, counted as heat. Transmission friction and the off-optimum motor current both become heat inside a sealed joint with no airflow — and what a motor can sustain is thermally bound. Sustained joint torque is a thermal spec, and the transmission spends the budget from both ends. Lost watts cost performance, not just charge.
- Payload → joint torque density. Every kilogram of gearbox in a knee is payload spent before the robot picks anything up.
- Durability → gear-tooth fatigue and impact tolerance. Motors don't mind footfalls. Gears do.
- Safety → backdrivability — how freely the joint can be pushed from outside — and reflected inertia (motor inertia × the square of the ratio: what a collision actually feels). A joint that yields mechanically is safe with the software crashed and the power off.
- Cost → precision part count: carrier, planet bearings, and a cross-roller output bearing that is routinely the most expensive part in the box — bought, pressed, preloaded, and replaced thirty times per robot.
Gearbox-limited
Every product-deciding problem, one component at the bottom of it. The gearbox between the motor and the world is where the humanoid's hardest constraints meet: humanoids are gearbox-limited. The transmissions the industry inherited were each optimized for a different question — that mismatch is the opening.
Next: the incumbents — harmonic, cycloidal, two-stage planetary — and which joints each one honestly suits.