One joint, two motors — and no motor in the joint

A serial ankle is the textbook design: one actuator on the pitch axis, and a second actuator — carried by the first — on the roll axis. It is simple to build and trivial to control, because each motor owns exactly one axis. The price is that at least one heavy actuator ends up at the very end of the leg, where its mass does the most damage. Every gram at the foot swings through the largest arc at the highest speed of any point on the robot, twice per stride.

A parallel ankle inverts the layout. The ankle itself becomes a passive two-axis joint — typically a universal joint — and both actuators move up the shank, connected to the foot through two independent kinematic chains: push rods on ball joints, or cranks and links. Push both chains the same way and the foot pitches; push them in opposition and it rolls. The two actuators share every load, and the mechanism is a closed kinematic chain — the same family of geometry as a flight-simulator platform, shrunk to fit behind a shin.

Why the actuators climbed the shin

The case for the parallel ankle is a case against distal mass. A design-optimization study presented at IEEE-RAS Humanoids 2025 lists the benefits that have driven adoption: heavy actuators move proximally, which trims the leg’s swing inertia and helps energy efficiency and agility; the closed chain adds structural rigidity; positioning errors average across the two chains instead of stacking; and because both actuators contribute to both axes, each one can be smaller than a motor that has to own an axis alone. Torque sharing also spreads a hard footstrike across two transmissions rather than sending it through one reducer’s teeth — a real concern when the alternative is a harmonic drive taking heel-strike shocks on its flexspline.

The same study puts a number on the payoff: re-optimizing an existing humanoid’s serial ankle as a rotary-parallel (RSU) mechanism reduced the design cost function — a weighted mix of the things ankle designers care about — by up to 41 percent. The lineage of the idea is older than the current humanoid wave: Agility Robotics’ Cassie, the ostrich-legged biped of 2017, kept its leg motors packed near the hip and drove the lower leg through linkages, and a 2025 control paper notes that several recently released humanoids, “inspired by the mechanical design of Cassie”, displace motors from the joints specifically to reduce leg inertia.

Who uses what

The adopter list reads like the current market. The Humanoids 2025 survey names Tesla’s Optimus, Unitree’s G1, PNDbotics’ Adam, Agility Robotics’ Digit, Fourier’s N1, TUM’s LOLA, Pandora and PAL Robotics’ Kangaroo as parallel-ankle designs. On Optimus the implementation is visible in Tesla’s own footage: active ankle pitch and roll driven by two of the company’s roller-screw linear actuators, plus a passive articulated toe section. Unitree’s G1 packs six degrees of freedom into each leg of a roughly 35 kg robot — a mass budget that a motor hanging on the foot would wreck. And Digit inherits the Cassie philosophy directly: actuators kept high, lower leg driven through closed-chain linkages.

SPU or RSU: the linear and rotary flavors

Within the parallel family there are two dominant layouts, named after their joint sequences. The SPU chain (spherical–prismatic–universal) uses two linear actuators as the legs of the mechanism — this is the Optimus pattern, and it suits builders who already have a roller-screw supply chain. The RSU chain (revolute–spherical–universal) uses rotary actuators turning cranks, with connecting links down to the foot — a natural fit for companies whose entire actuator portfolio is rotary modules. The Humanoids 2025 framework treats the choice as a genuine optimization problem: for the robot it studied, the optimized RSU beat both the serial baseline and the SPU alternative on the combined cost of mass, torque and workspace fit.

The control bill for the mechanical bargain

Parallel ankles are not free. What the mechanism saves in mass it charges back in mathematics, and in a few places where serial designs stay ahead.

• Loop-closure constraints. A closed chain cannot be simulated or controlled with the cheap recursive algorithms that serial robots enjoy. A 2025 paper on differential actuation models notes that these constraints “greatly increase computational cost”, which is why most whole-body controllers approximate the ankle as a virtual serial joint and translate torques afterwards.
• Simulation fidelity for learning. Reinforcement-learning pipelines meet the same problem at scale: work such as LiPS (2025) exists precisely to make large-scale RL training handle humanoids with parallel-series structures instead of pretending every joint is serial.
• Coupled, configuration-dependent behavior. The mapping from actuator effort to foot torque changes with ankle angle, and pitch and roll authority trade off against each other near the edges of the workspace — the reason academic kinematics treatments of humanoid parallel ankles run to full journal papers.
• More parts in the load path. Rod ends, ball joints and cranks add compliance, tolerance stack-up and wear points that a motor bolted on the axis does not have.

The counter-move: Unitree’s cable-driven H2

The most interesting recent data point pushes the same logic through a different mechanism. Unitree’s 31-DoF H2, detailed in late 2026 reporting, abandons the linkage-parallel ankle its predecessor used and instead runs what Humanoids Daily calls a remote-actuated “quasi-serial” ankle: the joint behaves like a serial two-axis ankle, but its actuators sit just below the knee and pull steel cables, tendon-style. The waist went the other way in the same redesign — from parallel mechanisms to a compact serial actuator stack. Both choices chase the same target the parallel ankle does: keep the heavy parts close to the body, minimize reflected inertia at the extremity — while trying to buy back the simple kinematics of a serial joint. Whether robot-scale cable transmissions can match linkage lifetimes is now one of the more consequential open questions in lower-leg design, and it echoes the tendon-wear debate already running in humanoid hands.

Either way, the direction of travel is one-way. Whatever transmission wins — push rod, crank or cable — no serious biped is putting a motor back in the foot.

FAQ

Sources