Valerant Uses World-Model Exploration to Generate Navigable 3D Game Maps

Summary

World Action Models combine predictive world modeling with action generation so anticipated future states can guide an agent. The paper argues that applying this idea to games is harder than applying it to robotics or autonomous driving: those domains have a persistent physical world, whereas a game must instantiate its own navigable space and, in 3D, explicit geometry. Existing game approaches generally operate in 2D visual-observation space and do not construct persistent 3D geometry. Valerant is a training-free framework that turns a pretrained action-conditioned world model into a World Action Model for exploration and map construction. It combines predictive visual rollouts with SLAM-based spatial reconstruction and exploration-driven action selection. Starting from a single image, the system progressively builds a persistent 3D game map that supports navigation and interaction. The work extends world-model-based interaction beyond 2D visual simulation and proposes automated map creation as a way to reduce manual effort in 3D game development.