humanoidsdata.com

Search

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

← All glossary terms

Hardware & control

Control frequency

Control frequency is the rate, usually expressed in hertz, at which a robot control loop reads state, updates its controller or policy, and issues commands. A robot can contain several nested loops at different rates, so policy frequency, actuator-control frequency, sensor sampling rate, and dataset frame rate should not be assumed to be identical.

Also known as: control rate, policy update rate, action frequency

Updated

One robot can run several clocks

A motor-current loop may run much faster than a joint controller, while a large learned policy updates only a few times per second and sends targets to those lower layers. Camera, tactile and logging streams can each use another rate.

The ros2_control controller manager defines its update rate as the frequency of the loop that reads hardware state, updates controllers and writes commands. Individual asynchronous controllers can run at different rates, which is why “the robot runs at 100 Hz” is incomplete without naming the loop.

Frequency changes the meaning of a command sequence

A Cartesian delta applied at 5 Hz produces a different velocity from the same delta applied at 20 Hz. Position targets may be less directly rate-dependent, but a slower loop observes disturbances later and offers fewer opportunities to correct.

RT-1 produced closed-loop actions at 3 Hz, while DROID recorded state and actions at 15 Hz. Neither number describes every inner hardware loop. It identifies the temporal resolution of the policy or released training sequence.

Nominal rate is not measured timing

A configured 100 Hz loop has a 10 ms target period, but computation overruns, scheduling jitter or dropped samples can reduce the achieved rate. Dataset metadata should include timestamps rather than relying only on a nominal frequency.

Documentation should distinguish capture, policy, command and actuator rates; state whether streams were resampled; and report timing variation where available. Cross-dataset training must account for those differences before action chunks or deltas are treated as comparable.

Sources