Source note

World Action Models: A Survey

This survey defines World Action Models as embodied predictive-action models that expose predicted futures to control. It maps WAMs against video generators, world models, and Vision-Language-Action policies, then explains the design tradeoffs that shape current systems.

  • WAM research has split across video-generation models, language-based action models, and vision-language robot policies, which makes it hard to compare methods.
  • The problem matters because robot control needs predictions that are useful for action, with limits on compute, memory, latency, and action-label cost.
  • The paper focuses on what a model must predict for control, rather than treating every future-prediction model as the same kind of robot model.
  • The survey first separates broad world models, video generation models, action-grounded video world models, Vision-Language-Action policies, and WAMs.
  • It organizes methods by what they generate: rendered futures, latent futures, or action reasoning without video generation.
  • It also decomposes methods along 4 axes: predictive substrate, backbone, action coupling, and deployment regime.
  • It uses those axes to compare interactability, causality, persistence, physical plausibility, generalization, data needs, evaluation, and open challenges.
  • The excerpt reports no benchmark scores, dataset metrics, or quantitative comparisons.
  • It claims a 2-view taxonomy: generation target and model anatomy.
  • It identifies 3 output regimes for WAMs: rendered futures, latent futures, and video-generation-free action reasoning.
  • It names 4 main design axes: predictive substrate, backbone, action coupling, and deployment regime.
  • Its main claim is that WAMs trade future-detail richness against compute, memory, latency, and action-label cost, with current work moving toward predicting less while keeping control-relevant information.