Game-Guided Skill Discovery through Self-Play for Playable Agent Control
Seungeun Rho, Jeonghwan Kim, Xue Bin Peng, Sehoon Ha
Georgia Institute of Technology · Simon Fraser University · NVIDIA
Published September 30, 2026 · on the feed October 1, 2026 · ▲ 5 · arXiv 2609.40137
The abstract
We present Game-Guided Skill Discovery (GGSD), a framework that uses self-play in games to discover motor skills that are directly playable by humans. Playable skills provide a compact abstraction for controlling embodied agents through a small set of learned behaviors rather than low-level actions. To be effective, these skills should be semantically distinct, interpretable, and expressive; properties that existing unsupervised skill-discovery methods often fail to achieve simultaneously. GGSD achieves these desiderata by grounding skill discovery in competitive gameplay. A hierarchical agent competes against its past selves, with a high-level policy selecting from a small discrete skill set and a skill-conditioned low-level policy learning the corresponding behaviors. After training, a human can replace the high-level policy and directly control the agent through the same discrete skills. Despite the small number of high-level actions, skill transitions give rise to emergent combo behaviors, expanding expressivity beyond individual primitives. Across Ant, Franka-arm, and Unitree G1 environments, we show that GGSD produces human-playable skills that humans can compose to solve unseen tasks, such as Maze and CubePush, without additional training. An interactive demo is available at https://ggsd-demo.github.io.
The Brief
2:09A short podcast summary of the paper. Written and voiced by AI.
Transcript
Game-Guided Skill Discovery through Self-Play for Playable Agent Control. By Seungeun Rho and three others, from the Georgia Institute of Technology, Simon Fraser University, and one other institution. Published on the thirtieth of September. No code repository is linked to it.
Picture controlling a robot the way you play a game: not through low-level actions, but by pressing one of a handful of buttons, each of which triggers a learned behaviour. Playable skills of that kind are a compact abstraction for controlling embodied agents. To work, though, the skills need three properties at once. They should be semantically distinct, interpretable, and expressive. Existing unsupervised skill-discovery methods often fail to deliver all three simultaneously.
Game-Guided Skill Discovery, GGSD, gets there by grounding skill discovery in competitive gameplay. The skills are discovered through self-play in games, and the aim from the start is skills a human can directly play.
The agent is hierarchical. A high-level policy chooses from a small, discrete set of skills, and a low-level policy, conditioned on the chosen skill, learns the behaviour that goes with it. This agent competes against its own past selves. After training, the high-level policy can be removed and a human put in its place, controlling the agent through exactly the same discrete skills. The small number of high-level actions might sound limiting. In practice, transitions between skills give rise to emergent combo behaviours, which extends what the agent can express beyond the individual primitives.
Experiments cover three environments: Ant, a Franka arm, and the Unitree G1. Across them, GGSD produces skills that humans can play, and that humans can compose to solve tasks they were never trained on, such as Maze and CubePush, without any additional training. An interactive demo is available online.
Instead of asking a discovery objective to stumble on skills that happen to be usable, GGSD lets a game decide what to discover, and holds the authors' three properties up as the result of that grounding. The human then inherits the same few buttons the high-level policy used, and the combos between them supply range that a short button list would otherwise lack. Playability is not added after training. It is the shape the agent was trained in.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a humanoid with countless ways to move and no reason to pick any.
You want people to control a robot like a game character: a handful of buttons, each triggering a learned behaviour, instead of setting every joint by hand.
The road ahead
- 1 · The problem ~7 min To steer a robot with a few buttons, its learned behaviours must be distinct, interpretable and expressive at once, and current methods miss that.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Let the robot compete in simple games against its past selves, so useful skills emerge, then hand the skill buttons to a human.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Set up a game, train a two-level agent against a pool of its past selves, swap in a human, then compose skills on new tasks.
- 6 · The experiment ~6 min You’ll guess: When humans took skills learned in games to new tasks like Maze and CubePush, how much extra policy training did those tasks need?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.