Open Models Almanac
Methodsmodel architecturechannel mixermixture of experts

specific method · filed under model architecture

Action expert

An action expert is one of two sets of weights in a single transformer, where each token is routed to one of the experts.1 Images and language prompts are routed to the larger VLM backbone, while inputs not seen during VLM pre-training are routed to the action expert, and the experts interact only through the self-attention layers.2,3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. implemented as a single transformer with two sets of weights (also known as experts [45]), where each token is routed to one of the experts — π0\pi_{0}: A Vision-Language-Action Flow Model for General Robot Control Thanks: Physical Intelligence, San Francisco, California, USA.…
  2. The images and language prompt, , are routed to the larger VLM backbone, which we initialize from PaliGemma. The inputs not seen during VLM pre-training, , are routed to the action expert — π0\pi_{0}: A Vision-Language-Action Flow Model for General Robot Control Thanks: Physical Intelligence, San Francisco, California, USA.…
  3. Since the experts interact only in the self-attention layers — π0\pi_{0}: A Vision-Language-Action Flow Model for General Robot Control Thanks: Physical Intelligence, San Francisco, California, USA.…

Also called Transformer action expert.

sources
2
models
2
labs adopt it
2
strongest
core

Documented in

Hugging Face
Physical Intelligence

Evidence

2 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
core 2
↑