humanoidsdata.com

Search

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

← All glossary terms

Hardware & control

Control latency

Control latency is the elapsed time between a relevant event or observation and the physical effect of the resulting robot command. It can include sensor exposure and transport, preprocessing, policy inference, network transfer, controller scheduling, actuator response, and communication back to a human operator. The measured boundary must be stated because each component has a different latency.

Also known as: robot control latency, control delay

Updated

Latency is a duration, not a rate

Control frequency says how often a loop aims to update. Latency says how long information or a command takes to travel through that loop. A 100 Hz scheduler can still act on an observation that is 80 ms old, and a low-frequency policy can have short computation latency between its scheduled updates.

The boundary matters. Inference latency measures model computation. Command latency may run from policy output to controller receipt. Observation-to-actuation latency includes sensing, processing and physical response. Teleoperation also has feedback latency from robot sensors to the operator.

Delay changes the state at execution

RT-1 treated inference time as part of a real-time control budget. As models, cameras or networks become slower, the world can move between observation and action.

Delay-Aware Diffusion Policy describes this observation-execution gap explicitly: an action selected for the sensed state can arrive after that state has changed. Asynchronous inference can prevent idle pauses, but it does not make the underlying observation fresh.

Average latency hides dangerous variation

Jitter changes delay from one cycle to the next. Occasional high-percentile latency or packet loss can be more harmful than a stable average because controllers and operators cannot compensate consistently. Teleoperation research also treats time-varying delay as a stability problem.

A dataset or system report should state timestamp locations, one-way or round-trip measurement, mean and percentile latency, jitter, clock synchronisation and each known component. Operator input, sent command, controller receipt and measured motion should remain separate when the collection system can record them.

Sources