Game-Guided Skill Discovery through Self-Play for Playable Agent Control
Self-play is used here to turn a small menu of learned motor skills into direct human controls for embodied agents.
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 movements. After training, a human can take over that high-level role and play the agent through the same skill set. The authors report results across Ant, Franka-arm, and Unitree G1 environments, including unseen Maze and CubePush tasks without additional training. Skill transitions can combine into emergent behaviors beyond the individual primitives. ArXiv · AI/CL/LG's 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 movements. After training, a human can take over that high-level role and play the agent through the same skill set. The authors report results across Ant, Franka-arm, and Unitree G1 environments, including unseen Maze and CubePush tasks without additional training. Skill transitions can combine into emergent behaviors beyond the individual primitives. ArXiv · AI/CL/LG's note
score 4