Source-linked AI summary
Valerant: An Automatic Navigable Game Map Generator via Action-Conditioned World Model Exploration
Yiran Qiao, Feng Wang, Jing Ma
TL;DR
Games lack the external persistent 3D substrate available in robotics, while existing game-oriented WAM approaches mainly operate in 2D visual space. VALERANT addresses this gap with a training-free system that combines action-conditioned rollouts, SLAM reconstruction, and exploration-based selection to build a persistent 3D map from one image; experiments report reliable autonomous exploration and incremental map construction under geometric drift and visual hallucinations.
Problem
General-purpose World Action Models remain largely unexplored in games, where visual rollouts do not by themselves instantiate the persistent navigable 3D geometry required by playable environments.
Method
VALERANT converts a pretrained action-conditioned world model into a WAM by evaluating candidate rollouts, reconstructing their geometry with visual SLAM, and committing selected observations to a persistent point-cloud map.
Results
VALERANT achieves autonomous collision-free exploration and incremental 3D game-map construction from a single image, supported by robust floor estimation, collision detection, and dead-pocket rewind.
Takeaways & Limitations
The framework provides a training-free approach for extending world-model interaction beyond 2D visual simulation toward navigable 3D game-map authoring.
Takeaways & Limitations
VALERANT remains subject to visual hallucinations and monocular SLAM pose and scale drift, motivating robust exploration mechanisms for geometrically consistent and collision-aware behavior.
Abstract
from arXiv · showhide
World Action Models (WAMs) couple predictive world modeling with action generation, allowing anticipated future states to guide agent behavior. Although WAMs are rapidly advancing embodied AI, general-purpose counterparts remain largely unexplored in games. Existing game-oriented approaches often combine action-conditioned world models with external policies and reward functions to realize WAM-like decision-making, yet they operate mainly in 2D visual observation space and do not instantiate persistent 3D geometry. Extending this paradigm to 3D games introduces a distinct challenge. In autonomous driving and robotics, the physical environment exists independently of the model, providing a persistent 3D world in which selected actions can be executed. Games have no such external substrate; the virtual world itself must be instantiated. Most playable games require a persistent and navigable space, while 3D games additionally require explicit geometry that supports movement and interaction. Action-conditioned video rollouts provide visual observations but not this spatial representation. We present \textsc{Valerant}, a training-free framework that transforms a pretrained action-conditioned world model into a WAM for exploring and constructing 3D game maps. By coupling predictive visual rollouts with SLAM-based spatial reconstruction and exploration-driven action selection, \textsc{Valerant} progressively transforms a single image into a persistent 3D game map. This framework extends WAM-based interaction beyond 2D visual simulation and offers a new approach to reducing manual effort in 3D game-map creation.
1 INTRODUCTION
VALERANT addresses the limited use of general-purpose World Action Models in games by converting a pretrained action-conditioned world model into a training-free system for constructing navigable 3D maps. It combines counterfactual visual rollouts, SLAM-based geometry reconstruction, and exploration-driven action selection to expand a single image into persistent game-map geometry.
- Motivation: World Action Models couple predictive world modeling with action generation or selection so anticipated futures inform an agent’s actions.This extends world modeling beyond simulating outcomes under externally specified actions.
- Motivation: Games require explicit persistent geometry for navigation and interaction, but action-conditioned video rollouts primarily provide visual observations rather than spatial representations.This creates a distinct challenge compared with physical environments, where the world already exists independently of the model.
- Approach: VALERANT transforms a single input image into a navigable 3D environment without additional training by coupling a pretrained action-conditioned world model with exploration-driven action selection.The framework queries candidate actions and evaluates their predicted futures before executing a selected action.
- Approach: Alternative world-model rollouts are converted into a persistent 3D map through SLAM-based reconstruction and integration of selected observations.The procedure bridges visual world simulation and the explicit spatial instantiation required by games.
- Implication: VALERANT proposes expanding a single visual concept into navigable 3D geometry to reduce reliance on labor-intensive game-map construction workflows.The stated contribution concerns automatic map authoring rather than only visual simulation.
2 RELATED WORK
Related work spans latent world models, game-specific World Action Models, visual SLAM, and counterfactual reasoning. Together, these lines provide the predictive, spatial, and alternative-action foundations relevant to VALERANT.
- World Models: Early world models learned action-conditioned transitions in compact latent spaces, while later methods improved scalability, robustness, and generality through imagined trajectories.These developments support learning across increasingly diverse control tasks.
- World Action Models: Game-oriented WAM research includes WHAM, which jointly models gameplay visuals and controller actions for gameplay ideation.This represents a game-specific joint WAM formulation.
- Visual SLAM: SLAM estimates an agent’s trajectory while incrementally constructing a spatial representation from sequential observations.Classical systems use geometric correspondences, bundle adjustment, and loop closure, while learning-based systems add learned components.
- Counterfactual Reasoning: Counterfactual reasoning considers how the future would change under an alternative action while holding preceding context fixed.World models make mutually exclusive candidate futures practically explorable before an action is executed.
3 METHOD
VALERANT converts a frozen action-conditioned video world model into a closed-loop explorer by branching candidate actions, reconstructing their predicted geometry with SLAM, and committing selected trajectories into a persistent 3D map. Robust floor normalization, collision checks, and checkpoint rewinding address geometric drift, hallucinated rollouts, and locally trapped exploration.
- World model and SLAM: Visual SLAM recovers camera trajectories and registers each generated rollout’s observed geometry in a common 3D coordinate system.The resulting representation supports persistent map construction and geometric candidate evaluation.
- World model and SLAM: The pretrained video world model remains frozen while VALERANT applies SLAM and exploration policy entirely at inference time.This converts passive action-conditioned visual predictions into closed-loop exploration without additional training.
- Counterfactual exploration: At each step, VALERANT evaluates five candidate action sequences: three forward-moving actions and two in-place rotations.All candidates branch from the same pre-action visual context, making their generated futures directly comparable.
- Counterfactual exploration: Candidate branches are scored by exploration coverage and forward progress, while collision-violating branches are rejected before one action is replayed and committed.Unsuccessful branches are discarded without modifying the persistent world state, and selected observations and geometry are fused into the next context and global map.
- Robustness scope: VALERANT must maintain geometrically consistent, collision-aware progress despite visual hallucinations and monocular SLAM pose and scale drift.These errors can produce unreliable geometry and unsafe exploration decisions.
- Collision-aware exploration: A locally adaptive floor estimator follows gradual monocular-SLAM drift while preventing non-ground surfaces from becoming the reference plane.Floor-relative normalization keeps ground and body-level occupancy bands aligned despite inconsistent global scale or elevation.
- Collision-aware exploration: Point-cloud collision detection tests reconstructed geometry against a normalized cylindrical body volume along each candidate trajectory.The method uses floor-relative height bands, suppresses isolated reconstruction noise, excludes ground and overhead structures, and exempts rotation-only candidates.
- Dead-pocket rewind: Checkpoint-based rewind restores synchronized WM and SLAM states when receding-horizon exploration stalls in a dead pocket.Previously traversed cells are retained as failed regions, steering subsequent candidates toward alternative routes while preserving visual and geometric synchronization.
4 EXPERIMENTS
VALERANT is evaluated as an automatic single-image map-generation system, with qualitative results showing autonomous exploration and persistent 3D reconstruction. Ablations further show that three robustness modules address distinct exploration failures.
- Experiment settings: Experiments use VALORANT screenshots as single-view inputs and compare VALERANT with HY-World 2.0 on the same initial frames.The comparison involves methods that both generate navigable 3D worlds from a single image, despite differing technically.
- Experiment settings: Table 1 reports quantitative comparisons using AS for aesthetic score and UP for user preference.The supplied table caption identifies the metrics but does not provide their numerical values.
- Experiment settings: Evaluation uses Authenticity Score and User Preference because the open-ended task lacks a unique ground-truth map.Authenticity Score is assessed by a VLM, while User Preference measures pairwise human preferences.
- Main results: Across four scenes, VALERANT automatically explores from the first frame without manually specified trajectories or intermediate intervention.Generated frames depict movement through corridors, turns, and open regions.
- Main results: The system avoids collisions while visual SLAM incrementally builds a persistent point-cloud map with substantial coverage along executed trajectories.The qualitative results jointly demonstrate autonomous collision-free exploration and 3D map construction from one image.
- Ablation study: RF, COL, and DP address complementary failures from floor-estimation drift, wall penetration, and local exploration traps, respectively.Removing each module causes a corresponding qualitative failure, while restoring it improves traversal, free-space compliance, or route selection.
5 CONCLUSION
VALERANT constructs navigable 3D game maps from a single image without training. It couples a frozen action-conditioned video world model with visual SLAM and exploration, while robustness modules support autonomous exploration; future work targets richer editable assets.
- VALERANT is a training-free framework that constructs navigable 3D game maps from a single image.
- The framework couples a frozen action-conditioned video world model with visual SLAM and an exploration policy to integrate selected observations into a persistent point-cloud map.
- Robust floor estimation, collision detection, and dead-pocket rewind support reliable autonomous exploration under geometric drift and visual hallucinations.
- Future work will improve geometric consistency and convert reconstructed point clouds into richer, editable game assets.