Towards Generalist Embodied AI: A Survey on World Models for VLA Agents

Vision-language-action (VLA) models represent a critical milestone toward generalist embodied AI. However, they often struggle to precisely capture physical dynamics, verify plan executability, and overcome data scarcity. To address these limitations, world models are introduced as future predictors that incorporate rich physical priors to effectively guide VLA agents toward physically grounded action generation. While this paradigm is emerging as a pivotal research direction, the field currently lacks a systematic overview. Therefore, we present the first survey specifically dedicated to world models for VLA agents. We propose a unified taxonomy that organizes existing methods into four paradigms: world planner, world action model, world synthesizer, and world simulator, and review their respective evolutionary trajectories. Furthermore, we summarize the ecosystem of foundation models, benchmarks, and evaluation metrics to support future research. Finally, we highlight key challenges and promising directions to facilitate the advancement of world models for generalist VLA agents.