Open Models Almanac
Methods

general family · not yet filed

Flow matching

Flow matching is a training objective that represents an action distribution by learning a vector field that denoises actions.1 At each step, action tokens take the partially denoised actions from the previous step as input and predict the flow matching vector field toward the clean actions.2,3

Trade-off It gives a more expressive representation of continuous action chunks than discrete tokenization.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. represents the action distribution via flow matching — π0.5\pi_{0.5}: a Vision-Language-Action Model with Open-World Generalization
  2. the tokens corresponding to actions receive the partially denoised actions from the previous step of flow matching as input, and output the flow matching vector field — π0.5\pi_{0.5}: a Vision-Language-Action Model with Open-World Generalization
  3. is trained to output the vector field from the VLM features and the noisy actions — SmolVLA: A vision-language-action model for affordable and efficient robotics
  4. providing a more expressive representation over continuous-valued action chunks — π0.5\pi_{0.5}: a Vision-Language-Action Model with Open-World Generalization

Also called Flow matching training objective, flow matching objective, flow-matching loss, Flow-matching action generation, Flow-matching trajectory decoding, Flow matching for action generation.

sources
8
models
5
labs adopt it
4
strongest
core

Documented in

Hugging Face
Physical Intelligence

Evidence

10 quotes from 8 sources, one card per model, each opening with its strongest.

How sources treat it
used 4default 1core 5

Xiaomi-Robotics-1

Xiaomiadopted2 sources

the DiT generates the action chunk via flow-matching [49]

coremodel architecturein Xiaomi-Robotics-1Xiaomi
1 more from this model

train the diffusion transformer (DiT) from scratch on the entire robot trajectory data with a flow-matching loss

usedtraining objectivein Xiaomi-Robotics-0Xiaomi Robotics

π0.5

Physical Intelligenceadopted2 sources

This stage jointly trains with next-token prediction, to preserve text prediction capabilities, and flow matching for the action expert

usedunclearin π0.5
2 more from this model

Training objective: flow matching

coretraining objectivein π₀.₅Hugging Face

is trained with the flow matching objective

usedtraining objectivein π0.5
↑