One-line meaning
A policy that generates actions by iteratively denoising a sampled action sequence conditioned on observations.
Nuance
Sampling cost, conditioning, prediction horizon, and execution horizon determine runtime behavior; it is not synonymous with VLA.