Action Expert

The Action Expert is the part of an autonomous-driving AI that decides "how to move now." If you think of the AI as having a part that perceives and understands the situation like eyes and a brain, and a part that actually turns the wheel and hits the brakes like hands and feet, the Action Expert corresponds to the hands and feet.

Large autonomous-driving models typically combine a component that understands and reasons over camera footage with a component that draws out a concrete driving trajectory based on that reasoning. The Action Expert is dedicated to the latter role, using a technique called diffusion to gradually refine a rough, blurry candidate trajectory into a final path of movement. Although it makes up only a small portion of the overall model size, it serves as the final gateway that turns judgment into actual movement.

In the article, it's described as "a total of 34 billion parameters combining the 32-billion-parameter Cosmos 3 Super Reasoner with the 2-billion-parameter, diffusion-based Action Expert." A common misunderstanding is that the Action Expert is not a standalone, complete autonomous-driving AI on its own. It's a component within a larger structure that only completes actual driving decisions when paired with the larger model that understands and reasons about the situation.