World Models and Vision-Language-Action Systems for Robotic Intelligence: Predictive Simulation, Policy Evaluation, and Whole-Body Control

Abstract

World models seek to represent how an environment evolves and thereby allow an intelligent system to predict the consequences of possible actions before executing them. Although latent world models have long been studied in reinforcement learning, advances in generative video, multimodal perception, and vision-language-action modeling have expanded the concept toward realistic physical simulation and robot control. This review examines world models and vision-language-action systems for robotics through September 2026. Early latent-dynamics methods demonstrated that agents could learn policies using imagined trajectories, while Dreamer-style systems progressively improved sample-efficient control through learned representations of environmental dynamics. Diffusion Policy introduced expressive diffusion-based action generation for visuomotor behavior. RT-1, RT-2, Open X-Embodiment, and Octo demonstrated increasingly general robotic policies trained across heterogeneous datasets and tasks. Generative video systems subsequently offered richer visual predictions of object and environment dynamics. By 2025–2026, this line of work increasingly converged with world modeling. PhysWorld explored robot learning from physical video-generation models. WorldGym used an action-conditioned video model as an environment in which robotic policies could be evaluated through simulated rollouts and vision-language-based rewards. Sequential World Models extended model-based reinforcement learning to multi-robot cooperation through autoregressive agent-wise dynamics. WholeBodyVLA integrated vision-language-action learning with latent actions and reinforcement-learning control for humanoid loco-manipulation. Research on video generation for robotics and world-action models increasingly formalized the transition from predicting future observations toward using predicted futures to determine executable actions. The review identifies physical fidelity, causal correctness, action grounding, sim-to-real transfer, long-horizon error accumulation, and safety as central challenges.