A new preprint on arXiv proposes a unifying framework for two prominent families of robot learning models: vision-language-action (VLA) models and world-action models (WAMs). Both approaches combine multimodal perception with action generation, but they have developed along separate lines. The paper's title points to a "model-to-policy formulation" as the basis for bringing them together.

The abstract explains that VLA models and WAMs have emerged as promising routes to robot learning and control, yet it stops short of detailing exactly how they differ or how the proposed formulation resolves those differences. The visible portion of the abstract only sets up the motivation: these models integrate perception and action, but their divergence motivates a unified treatment.

Because the abstract is truncated, the specifics of the proposed formulation are not available from this source. What is clear is that the authors see value in a common framework that can span both model types, potentially simplifying how embodied foundation models are designed and compared. As a preprint, the work has not yet undergone peer review.