Policy Inference Frequency

How many times per second a learned policy is called to compute a new action, often lower than the control frequency below it.

Policy inference frequency is how many times per second a neural-network policy runs a forward pass to produce a new action (or a new chunk of actions), measured in Hz. It is not the same as control frequency: control frequency is how often the actuators receive a new command, and joint control loops commonly run at hundreds to a few thousand Hz. Large models are slow to run, so they typically predict a whole block of actions at once (action chunking) and execute that block step by step between inferences, so inference frequency ends up far lower than control frequency. π0, for example, outputs 50 steps of actions at once; on a 50 Hz robot it executes 25 of those steps before inferring again, roughly 2 Hz, with a single inference taking about 73 milliseconds on an RTX 4090. The lower the inference frequency, the slower the reaction to sudden changes, and the more likely stutters appear at the seams between chunks — which is why asynchronous inference, real-time chunking, and fast-slow dual systems (e.g., Figure's Helix, with a 7–9 Hz slow system and a 200 Hz fast system) exist.

Exampleπ0 re-infers every 16 steps (0.8 seconds) of execution on 20 Hz-controlled UR5e and Franka arms, and every 25 steps (0.5 seconds) on robots controlled at 50 Hz.

Also called

Inference Frequency, Model Call Frequency

Related

Control Frequency · Inference Latency · Action Chunking · Asynchronous Inference · Real-Time Chunking · Dual-System Architecture (System 1 / System 2)

Sources

π0: A Vision-Language-Action Flow Model for General Robot Control (arXiv:2410.24164)
Figure: Helix（System 2 7–9 Hz / System 1 200 Hz）
Real-Time Execution of Action Chunking Flow Policies (arXiv:2506.07339)

As of

2025-06