Learning Controllable 3D Level Generators

Key insights

6 takeaways

01
Reinforcement learning combined with neural cellular automata enables the procedural generation of functional 3D voxel environments that satisfy complex constraints like gravity and pathfinding.See it on page 1
02
Generative agents successfully master the construction of intricate, stair-based geometries to manipulate path lengths and define spatial boundaries within levels.See it on page 1
03
Current models struggle with reliability in large-scale generation, specifically regarding the connection of vertically stacked portals and maintaining consistency over long path sequences.See it on page 6
04
The framework supports mixed-initiative design, allowing human creators to collaborate with algorithmic agents to refine level architecture.See it on page 9
05
Future development requires more sophisticated curriculum-based training strategies and nuanced 3D complexity metrics to improve output stability.See it on page 1
06
This methodology is designed to scale with player progress, offering a pathway to automate curriculum learning for player agents.See it on page 8

Summary

Procedural content generation via reinforcement learning offers a robust framework for creating complex, traversable 3D environments within voxel-based systems like Minecraft. By leveraging proximal policy optimization and neural cellular automata, these generative models successfully navigate functional constraints such as gravity and pathfinding to produce structured layouts including mazes and dungeons. The primary objective is to establish a controllable design process where agents learn to satisfy specific architectural requirements, such as maximizing maze diameter or ensuring logical door connectivity, thereby facilitating both co-creative design and automated curriculum learning for player agents.

Empirical testing reveals that these agents demonstrate significant proficiency in constructing intricate, stair-based geometries to extend path lengths and define spatial boundaries. However, the current implementation faces limitations regarding reliability, particularly when tasked with connecting vertically stacked portals or maintaining consistency across extremely long path sequences. These findings suggest that while the models effectively manage fundamental level metrics, achieving high-fidelity, large-scale generation requires further architectural refinement and the integration of more sophisticated curriculum-based training strategies to improve output stability.

Ultimately, the integration of reinforcement learning and quality diversity methods provides a scalable approach to generating adaptive game content that evolves alongside player progress. This methodology proves particularly effective for mixed-initiative design, where human creators and algorithmic agents collaborate to refine level architecture. Future advancements in this field will likely focus on developing more nuanced 3D complexity metrics and expanding these generative techniques beyond voxel-based environments, potentially broadening the application of controllable procedural generation across diverse game genres and development workflows.