humanoidsdata.com

Search

Search companies, datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Hardware & control

Underactuation

Underactuation is the condition in which a system’s available control inputs cannot independently produce every possible acceleration of its modelled degrees of freedom. It often arises because there are fewer independent actuators than degrees of freedom, but contact, actuator limits and the chosen model also determine whether a system is underactuated.

Also known as: underactuated system, underactuated robot

Updated

Actuator count is a useful shortcut, not the full definition

A mechanism with fewer independent actuators than modelled degrees of freedom is underactuated. The formal MIT definition asks whether the controls can produce every desired generalized acceleration at a particular state and time.

That distinction matters because underactuation depends on the model and constraints. A rigid two-joint arm with two motors may be fully actuated in joint space, while a model that also includes unactuated link flexibility is not. Torque limits can prevent accelerations that an unconstrained model would consider available.

Humanoid base motion is not directly actuated

A floating-base humanoid adds six coordinates for root translation and rotation, but it has no motor attached between its pelvis and the world. Joint torques influence that base only through inertia and external contacts. Walking is additionally constrained because a foot can push against the ground but cannot pull on it, and friction limits horizontal force.

Contacts can temporarily constrain otherwise free motion, yet those constraints change as feet lift, land or slip. Whole-body control therefore reasons about actuation and feasible contact forces together.

Underactuated does not mean uncontrollable

An underactuated robot cannot track every arbitrary state trajectory, but it may still reach and stabilise useful states by exploiting dynamics. Swinging up a pendulum, balancing a biped and redirecting momentum before a footstep are standard examples.

Data and models should state which coordinates are actuated, actuator and transmission limits, passive joints, contact assumptions and any external support. Treating simulator-only forces at a floating base as ordinary robot actions can make a policy appear feasible while bypassing the real underactuation.

Sources