The authors propose Game-Guided Skill Discovery (GGSD), which uses self-play in simple competitive games to learn a compact set of motor skills for an embodied agent. The agent employs a hierarchical policy: a high-level policy selects from a small discrete skill set, and a low-level policy executes the corresponding behaviors. Training involves competing against past selves, with a mutual-information objective to encourage distinct skills.
The learned skills are semantically distinct, interpretable, and directly playable by humans. After training, humans can control the agent by selecting skills via keyboard, and they can combine these skills to solve previously unseen tasks (Maze, CubePush) without any additional training. This work is interesting because it demonstrates that competitive game structure can guide unsupervised skill discovery to produce behaviors that are both intuitive for human control and reusable for new tasks, potentially enabling more natural interfaces for robot control.