Human-aligned AI is a critical component of co-creativity, enabling models to accurately interpret human intent and generate controllable outputs that align with design goals. This is particularly relevant in Procedural Content Generation via Reinforcement Learning (PCGRL), which aims to serve human designers.
However, existing systems often fail to exhibit human-centered behavior, limiting the practical utility of AI-driven generation tools in real-world workflows. This paper proposes VIPCGRL (Vision-Instruction PCGRL), a novel deep reinforcement learning framework that integrates three modalities: text, level, and sketches, to extend control and enhance human-likeness.
We introduce a shared embedding space trained via quadruple contrastive learning across modalities and human-AI styles, aligning the policy using an auxiliary reward based on embedding similarity. Experimental results demonstrate that VIPCGRL outperforms existing baselines in human-likeness, validated by both quantitative metrics and human evaluations. The code and dataset are available at GitHub.
Blogger's Review: The VIPCGRL framework effectively enhances the quality of procedural content generation through multimodal learning, showcasing significant improvements in human alignment. This innovative approach provides new insights for future co-creation tools, making it a noteworthy area for further exploration.