World-Action Models (WAMs) are emerging as a second major paradigm for robot foundation models alongside VLM-based Vision-Language-Action models (VLAs). WAMs leverage pretrained video or world-model backbones to bridge the language-to-action grounding gap that limits classical VLAs. The post surveys the modern WAM landscape across three design axes: paradigm (inverse dynamics, joint prediction, representation-only), action integration (default tokens, action-as-image, latent actions/plans), and architecture (hierarchical, monolithic, Mixture-of-Transformers). Key models covered include UniPi, GR-1, LingBot-VA, DreamZero, Cosmos Policy, Being-H0.7, and Fast-WAM. A qualitative experiment with Google's Veo 3.1 illustrates how frontier video models already encode useful robot manipulation priors without robotics-specific fine-tuning. Practical challenges include high training compute (WAMs can be ~7x more expensive than efficient VLA recipes), slow inference due to video denoising, and significant memory/engineering overhead. The author predicts WAMs will become the second major recipe for robot foundation models, with the likely winner being a hybrid of VLA and WAM approaches.
