World Action Models (WAMs) combine predictive world modeling with action generation, allowing anticipated future states to steer agent behavior. While WAMs are advancing quickly in embodied AI, their general‑purpose use in games remains largely unexplored. Existing game‑oriented methods typically fuse action‑conditioned world models with external policies or reward functions to achieve WAM‑like decision making, but they operate mainly in 2D visual observation spaces and lack persistent 3D geometry.
In autonomous driving and robotics, the physical environment exists independently of the model, offering a stable 3D world where actions can be executed. Games have no such external substrate; the virtual world itself must be instantiated. Most playable games require a persistent, navigable space, and 3D games additionally need explicit geometry to support movement and interaction. Action‑conditioned video rollouts provide visual observations but not spatial representations.
We introduce Valerant, a training‑free framework that converts a pretrained action‑conditioned world model into a WAM for exploring and constructing 3D game maps. The key idea is to couple predictive visual rollouts with SLAM‑based spatial reconstruction and drive action selection through exploration, progressively turning a single image into a persistent 3D map. This extends WAM‑based interaction beyond 2D visual simulation and offers a new way to reduce manual effort in 3D game‑map creation.
Review