PlayCraft: Designing Engaging Games from Agent Play Experience
PlayCraft: Designing Engaging Games from Agent Play Experience
PlayCraft is an automated framework that uses vision-based agents to repeatedly playtest games and revises them based on playtest reports to improve learnability. The framework alternates between two phases: player agents observe the screen, take actions, and retain memory, then report progress, outcomes, decisions, and reflections after each session; designer agents estimate player ability from these reports, predict performance across different skill levels, and diagnose and patch gaps by comparing against target learning curves.
The paper presents an automated framework driven by vision-agent playtest reports to iterate level design, and compares learning gains and learning slopes across three agent configurations.
Source: arXiv Game AI · arxiv.org