Planetary Actuator for a Humanoid Knee

Which Transmission Belongs in a Knee

The lower limbs of a humanoid are one of the few places in robotics where a single-stage planetary gear beats the high-ratio alternatives. Strain-wave and cycloid drives reach far larger reductions in one compact stage, but they pay for it twice: high reflected inertia and poor backdriveability. Foot-strike impacts are difficult to absorb by mechanical yielding, which places greater demands on high-bandwidth control, and motor current is too decoupled from output torque to serve as a force sensor. A planetary stage trades much of that ratio away and gets both properties back — impacts reflect through to the motor, current becomes a usable torque proxy, and impedance control no longer has to fight the transmission.

Planetary gear sets are well-trodden ground in wind turbine, automotive and bicycle hub drivetrains. What makes the knee worth a dedicated model is that the gearbox can no longer be studied in isolation: with little torque multiplication downstream, the actuator needs a motor capable of delivering much of the knee torque itself.

A Simpler Gearbox, a Harder Motor

The actuator modelled here reaches the required torque density with a 24-slot, 20-pole machine using fractional-slot concentrated windings for a high copper fill factor. Motors of this type, asked for knee-level torque, routinely operate close to magnetic saturation — which rules out linear magnetic material models from the outset.

Three Coupled Domains

The magnetic field is described by the diffusion equation for the magnetic vector potential and solved by nonlinear, time-dependent finite element analysis: nonlinear because reluctivity depends on saturation, time-dependent to capture the eddy currents induced in the magnets and the laminated iron. That field model is two-way coupled to a voltage-driven three-phase electrical circuit and to a mechanical model of the rotor and planetary gear set, the latter kept as a comparatively simple analytical lumped-parameter model.

From Hours per Stride to Seconds

The complete model carries roughly 87,000 degrees of freedom, and an implicit solver needs several hours of wall-clock time per simulated stride. Model-order reduction closes that gap in two steps. A POD-Galerkin projection first maps the magnetic vector potential onto a small set of dominant field patterns. Hyperreduction with ECSW then targets what becomes the dominant cost once the model itself is small — evaluating the saturation nonlinearity — replacing the full mesh of roughly 173,000 elements with a weighted subset of only a few hundred. Together they bring a simulated stride from hours down to seconds.

Significance

The knee actuator is a clean illustration of a trade-off that recurs throughout mechatronic design. The backdriveability that makes a planetary stage the right choice for a leg is bought by pushing the difficulty upstream into the motor, which then has to be modelled in a regime — saturated, carrying eddy currents, electrically coupled to its drive — where simpler magnetic models stop being predictive. Complexity rarely disappears in engineering; it only changes places. A coupled model fast enough to run in seconds is what makes it possible to see exactly where it went.