Open Models Almanac
Methodsinference & servinginference scheduling

general family · filed under inference & serving

Asynchronous inference and execution

Asynchronous inference is an inference-scheduling method for robot policies that decouples action execution from policy inference, allowing predictions and control to run in parallel.1 While the next action chunk is being inferred, the robot keeps executing the remaining actions of the current chunk, so actions are always available for execution throughout inference.2,3

Trade-off It achieves comparable success rates to synchronous inference while substantially reducing task completion time.4

How this definition was written

Summarised by glm-5.3-flash from 2 sources; every sentence is checked against the quote it cites.

  1. the async mode decouples action execution and policy inference, allowing predictions and control to run in parallel — SmolVLA: A vision-language-action model for affordable and efficient robotics
  2. the robot continues executing the remaining actions of the current chunk while the next chunk is being inferred — Xiaomi-Robotics-0: An Open-Sourced Vision-Language-Action Model with Real-Time Execution
  3. there are always actions available for execution throughout the entire inference — Xiaomi-Robotics-0: An Open-Sourced Vision-Language-Action Model with Real-Time Execution
  4. both inference modes achieve comparable success rates across three real-world tasks. However, asynchronous inference demonstrates a substantial speed advantage — SmolVLA: A vision-language-action model for affordable and efficient robotics

Also called async scheduling, asynchronous (async) inference stack, asynchronous inference, Asynchronous inference for robot policies, Asynchronous inference scheduling, asynchronous execution.

sources
7
models
6
labs adopt it
6
strongest
core

Documented in

Evidence

9 quotes from 7 sources, one card per model, each opening with its strongest.

How sources treat it
optional 2used 4core 3

SmolVLA

Hugging Faceadopted2 sources

we propose an asynchronous inference stack that enables faster adaptation and responsiveness in real-world manipulation tasks.

coreunclearin SmolVLA
3 more from this model

Control loop keeps executing → inference happens in parallel (non-blocking).

coreinference servingin SmolVLAHugging Face

We directly assess the lack of adaptiveness of robot systems due to acting open-loop, and the presence of lags at runtime by decoupling action chunk prediction from action execution, developing an asynchronous (async) inference stack

usedinference servingin SmolVLAHugging Face

Finally, we evaluate SmolVLA under synchronous and asynchronous inference modes.

optionalinference servingin SmolVLAHugging Face

Further reading

Picked by hand, not extracted: where to read more, not evidence for anything on this page.

↑