VLAW: Iterative Co-Improvement of Vision-Language-Action Policy and World Model

ABSTRACT

The goal of this paper is to improve VLA performance through iterative online interaction. Since collecting real-world rollouts is expensive, we investigate whether a learned simulator—an action-conditioned video generation model—can generate rollout data. We propose VLAW, an iterative algorithm that uses real-world rollout data to improve the physical fidelity of a world model, which then generates synthetic data to improve the VLA policy. In experiments on a real robot, we achieve a 39.2% absolute success rate improvement over the base policy. We show that the world model learns to capture complex dynamics and failure cases, enabling the policy to progressively overcome execution errors.