The six-joint budget

A human leg, in the simplified models roboticists actually use, is about seven degrees of freedom: a ball-joint hip, a knee, and an ankle complex that pitches, rolls and twists. Robots cannot afford a copy. Every degree of freedom is a motor, a reducer, an encoder pair and a driver board, plus the mass and cable routing to feed them, so the leg is designed backwards from the question: what is the minimum set of joints that still balances and walks like a human?

The industry’s answer is six. The hip gets three orthogonal axes — yaw to steer the leg, roll to shift weight sideways, pitch to swing it forward — because balance is a three-dimensional problem and the hip is where the whole body’s weight is redirected. The knee gets exactly one axis, pitch, because that is all the human knee meaningfully contributes to gait. The ankle gets two, pitch and roll, so the foot can stay flat on ground that isn’t. Unitree’s G1 is the textbook case: 23 powered degrees of freedom on the base robot, six per leg, with the knee rated at 90 N·m (120 N·m on the EDU version) from low-inertia internal-rotor PMSM motors.

Cheaper or earlier designs cut the ankle first. Unitree’s own H1, the older and larger robot, runs five degrees of freedom per leg — the ankle pitches but cannot roll — and 19 in total. It walks fast anyway, but a pitch-only ankle forces the controller to handle every sideways disturbance at the hip, and the foot cannot conform to lateral slopes. The trend since 2024 has run the other way: ankles are gaining a second axis, and the interesting engineering question has become how to drive it.

Where the torque goes

The three leg joints are not equal customers. Motor maker CubeMars, whose actuators appear in a long list of research humanoids, publishes a selection hierarchy that matches what robot spec sheets imply: the hip dominates. It redirects the full body weight through multiple axes, sweeps roughly ±90 degrees in flexion, and has to survive squatting, jumping and landing, so CubeMars sizes hip motors for peak torque three to five times the rated figure. The knee is a single-axis hinge over about 0–135 degrees and gets away with two to four times rated torque, with the emphasis shifting to axial compactness and thermal stability, because the knee works continuously in every step. The ankle needs less torque than either but pays a different price: whatever drives it hangs at the end of the leg, where mass hurts most.

Published joint figures are still rare — most builders quote one headline number rather than a per-joint table. Unitree is the most transparent: the G1’s knee peaks at 90 N·m (120 on the EDU), while the H1’s M107 joint motor peaks at 360 N·m, the figure behind its most heavily loaded hip and knee axes. At component level, CubeMars’ flagship hip motor, the AK10-9 V3.0, is rated 18 N·m continuous with a 53 N·m peak. The spread in the chart below is the point: a heavy, fast biped needs several times the joint torque of a 35 kg research platform, and the actuator bill scales with it.

The ankle is the hard joint

The ankle’s problem is not torque, it is location. Every gram at the foot swings back and forth with every step, so distal mass costs energy and control bandwidth twice over. Mount a motor and reducer directly on the ankle axis — the serial layout — and you get simple kinematics and the worst possible mass distribution. The alternative, now spreading fast, is the parallel mechanism: both ankle actuators sit high on the shank, near the knee, and drive the foot through pushrods or crank linkages. Pitch and roll are produced differentially — rods moving together pitch the foot, rods moving in opposition roll it — so the two actuators share the load in both axes instead of each owning one.

A September 2025 design framework from researchers at the Italian Institute of Technology lists who has already made the switch: Tesla’s Optimus, Unitree’s G1, PNDbotics’ Adam, Agility’s Digit, Fourier’s N1, and research platforms LOLA, Pandora and Kangaroo all use parallel ankles. The same paper separates the two dominant recipes — SPU chains built on linear actuators, as in Optimus, and RSU crank-rocker chains built on rotary actuators — and reports that an optimised RSU ankle cut the design cost function by up to 41 percent against the serial design it replaced, and 14 percent against a conventionally engineered parallel version. The gains come from proximal actuator mass, torque sharing between the two chains, and stiffer, more precise output from the closed linkage.

The price is paid in software. A closed kinematic chain has no simple one-motor-per-axis mapping, so controllers and learning pipelines that assume a serial robot model need explicit differential-actuation models to command a parallel ankle properly — the subject of a dedicated 2025 paper on controlling humanoids with parallel mechanisms. That work is why simulation frameworks have been racing to add closed-chain support: the hardware trend got ahead of the tooling.

Twelve joints, most of the money

Follow the actuator count and the leg’s economic weight becomes obvious. On the G1, twelve of 23 joint motors are in the legs. On Optimus, analyst reconstructions of the 28-actuator body put six actuators in each leg, including the largest linear units in the robot — the ones rated to 8,000 N — on the knee and ankle. The leg joints are simultaneously the highest-torque, the most thermally stressed and the most safety-critical actuators on the machine, which is why they concentrate the expensive parts: the biggest motors, the highest-grade transmissions, and the force or torque sensing that keeps a 35–90 kg machine from falling over.

It also explains the architectural split we covered in the linear vs rotary explainer: builders who pay for planetary roller screws put them precisely where this page’s numbers point — knees and ankles, the high-load hinges — while hips, which need a wide, evenly loaded range of motion in three axes, stay rotary on almost every robot. The leg is not just where the torque goes. It is where the architecture argument is settled.

FAQ

Sources