New PCGRL framework generates playable and diverse complex platformer levels via inflated envelopes and forward curricula
PCGRL for Complex Platformers: Synthesizing Playable and Diverse Levels via Inflated Envelopes and Forward Curricula
Researchers propose a physics-aware, curriculum-guided PCGRL framework for generating solvable and diverse platformer levels on large canvases. The framework uses an "Inflated Volume Envelope" to embed continuous kinematic feasibility into the discrete state space of an MDP, and introduces intermediate state-transition feedback and multi-stage forward-expansion curriculum learning to alleviate sparse rewards. In the synthesis stage, it employs high-entropy stochastic inference combined with post-hoc parallel rejection filtering to improve diversity, and in multi-scale comparisons it outperforms traditional PCGRL baselines in convergence stability and scalability.
Source: arXiv Game AI · arxiv.org