Game-Guided Skill Discovery through Self-Play for Playable Agent Control
The paper’s hook is a self-play system that turns learned robot behaviors into a small human-playable control set.
GGSD trains a hierarchical agent against past versions of itself, with a high-level policy choosing discrete skills and a low-level policy learning the motions. After training, a person can take over that high-level role and control the agent through those same skills. The authors report results across Ant, Franka-arm, and Unitree G1 environments, including unseen Maze and CubePush tasks without extra training. Skill transitions can also combine into emergent behaviors beyond the individual primitives. HF Daily Papers' note
GGSD trains a hierarchical agent against past versions of itself, with a high-level policy choosing discrete skills and a low-level policy learning the motions. After training, a person can take over that high-level role and control the agent through those same skills. The authors report results across Ant, Franka-arm, and Unitree G1 environments, including unseen Maze and CubePush tasks without extra training. Skill transitions can also combine into emergent behaviors beyond the individual primitives. HF Daily Papers' note
score 4