Skill DiscoveryExpert
Game-Guided Skill Discovery through Self-Play for Playable Agent Control
This preprint introduces Game-Guided Skill Discovery (GGSD), a self-play framework that learns a small set of discrete motor skills that humans can directly use to control simulated agents. Across Ant, Franka arm, and Unitree G1 embodiments, GGSD produced semantically distinct skills and let users with brief practice solve unseen maze and cube-push tasks with 84–100% success. It matters because it points to compact, human-playable skill interfaces for high-degree-of-freedom robots without task-specific retraining.