Vision Language Actions using latent representation as action tokens

Vision Language Action (VLA) models use latent representations as action tokens, meaning actions are not hard-coded or discretized beforehand, but learned as compact latent vectors. These latents are generated by the model (often via an autoencoder or diffusion model) and treated like tokens in a sequence, enabling unified reasoning over perception, language, and control while improving generalization and scalability across tasks and embodiments.