Why makers fight for fewer actuator designs

Every distinct actuator is its own bill of materials. It has its own motor lamination, its own gear set, its own housing tooling, its own end-of-line test and its own row in the spares catalogue. It also needs its own motor model in simulation, which matters more than it used to now that most humanoid controllers are trained in simulation and then transferred to hardware. Cut the number of designs and each remaining design gets more volume, more test hours and better-tuned models.

Tesla made the point publicly at AI Day 2022. Its actuator team compared the torque, speed and energy each joint needed and fed them into an optimisation model, which grouped the robot’s 28 structural actuators into a small set of shared designs: three rotary and three linear. (Robotics 24/7’s write-up quotes seven types, but describes the same three-plus-three split.) The motivation wasn’t elegance. It was the car-industry habit of designing for volume from day one.

Research robots get there from the other end. The MIT Mini Cheetah, whose actuator design was later copied across dozens of legged robots and humanoid joint modules, used three identical, low-cost motors in each leg, each with a 6:1 gearbox. The team compared assembling them to Lego. One part, twelve times, and a broken joint becomes a cheap swap rather than a repair job.

What the robots actually use

Few humanoid makers publish an actuator count by type, so the picture has to be pieced together from talks, spec sheets and research papers. Two patterns show up anyway. Low-cost and open designs converge on two to four sizes. Full-size commercial robots need a wider torque range, from wrist joints well under 50 N·m to knees in the hundreds, and that range is what forces extra designs.

Two sizes, 22 joints: Berkeley Humanoid Lite

Berkeley Humanoid Lite is the cleanest public example of extreme commonality. The open-source robot uses just two actuators, both built around a 3D-printed cycloidal reducer chosen because it spreads load over many teeth and forgives the limited resolution of a desktop printer. The larger 6512 unit, around 20 N·m peak, drives the ten leg joints. The smaller 5010, around 12 N·m, handles the twelve arm and gripper joints. Both run on the same off-the-shelf motor driver board.

The payoff is a parts list a student can actually source. The team puts the 6512 at US $188 and the 5010 at US $136 when bought in the US, and the whole robot at US $4,312, or US $3,236 with parts sourced in China.

The torque ladder suppliers sell

Actuator suppliers have turned commonality into a product strategy. Instead of custom joints per robot, they offer a ladder of sizes and let the robot maker pick a rung for each joint. RobStride, whose quasi-direct-drive modules show up in many Chinese humanoids and research robots, sells seven RS-series models. Peak torque runs from 5.5 N·m on the RS05, aimed at wrists and grippers, to about 121 N·m on the RS04. The RS03 in the middle offers 60 N·m peak and 21 N·m rated from a 9:1 ratio in roughly 900 g.

Korea’s TPC Robotics launched a seven-product humanoid lineup in September 2026 that mirrors Tesla’s split almost exactly: three rotary actuators for joints such as the neck, elbow and hip, three linear actuators for heavily loaded areas such as the thigh and calf, and a finger joint module. Germany’s Schaeffler is going further. It wants its integrated planetary joint actuator to become what it calls the ABS pump of humanoid robotics, a standard module reused across brands and platforms, and it is proving it out in its own factories through a deal to deploy Hexagon’s AEON humanoid.

What commonality costs

Sharing a design means some joints get more actuator than they need. That’s harmless at the hip. It hurts at the wrist and ankle, where every extra gram sits at the end of a long lever, raises limb inertia and has to be carried by every joint upstream. It is one reason full-size robots end up with more designs than small ones: Unitree’s H1 spans roughly 59 N·m at the ankle to 360 N·m at the knee, a sixfold range no single actuator covers well.

Designers have a few tricks for stretching one size across more joints. Parallel linkages let two identical actuators share an ankle’s pitch and roll load, so neither has to be sized for the worst case alone; the H1-2 lists its ankle as 75 × 2 N·m for exactly this reason. Linkages and screws change the effective ratio without a new motor. And peak ratings flatter. A family chosen on peak torque can still overheat on a joint with a heavy continuous load, which is why RobStride’s own guidance tells engineers to size on the rated figure.

Questions worth asking a robot or actuator vendor

• How many distinct actuator part numbers does the robot carry, and which joints share them? Fewer means a shorter spares list for a fleet.
• Are the shared actuators sized on continuous torque at the hardest joint, or on peak torque? The answer tells you where thermal limits will bite first.
• Do all sizes use the same driver, bus and command format? Berkeley Humanoid Lite’s single driver board is as much a commonality win as its two gearboxes.
• Is the actuator family sold to other robot makers? Shared volume across OEMs is what pushes unit cost down, and what keeps parts available if a robot model is discontinued.

FAQ

Sources