GGSD turns five buttons into playable skills via 1v1 self-play: humans clear Maze and CubePush on Ant, Franka and G1 with no extra training

GGSD shifts the objective of skill discovery from 'distinguishable state distributions' to 'behaviors that win', using simple 1v1 game rules as lightweight guidance and self-play against a pool of historical checkpoints, yielding semantically clear and interpretable skills on high-degree-of-freedom embodiments. Existing unsupervised skill discovery mostly relies on mutual information or Wasserstein dependency measures, which only guarantee that skills are distinguishable; language descriptions, demonstrations or reference motions as guidance become costly in high-dimensional state spaces or data preparation. GGSD replaces that guidance with a few lines of game rules, letting behavioral structure emerge from what is needed to win. The paper trains on four games, AntSumo, AntFencing, FrankaAirHockey and G1Boxing, with 4,096 parallel environments and 70k updates each; qualitative visualizations show Ant learning turning, locomotion and pushing, and G1 learning locomotion, punching and a guard-like arm raise, with key behaviors such as locomotion and striking recurring across random seeds.

A small discrete skill set gains expressivity through state-dependent skill transitions that produce emergent combos without adding human inputs. Individual skills can be static or single-purpose in isolation, for example Franka skills that only move the end effector to a table region and stay there; transitions between skills instead generate rapid, purposeful puck-striking motions, and in G1 switching from skill 5 to skill 2 produces a backward lean while the reverse transition produces a strong forward punch. The paper visualizes this by fixing one skill for five seconds and overlaying motion snapshots, and the appendix interprets it as the state space containing a small number of behavioral regimes within which skill semantics stay predictable; the authors note that too many regimes would give one button too many meanings and make the interface hard to understand.

When skill semantic diversity is measured with a vision-language model, GGSD exceeds baselines including DIAYN, DADS and METRA across all embodiments. The metric captions skill rollout videos with a VLM, embeds the captions and computes pairwise distances, assigning zero distance to clips marked as having no semantic meaning, so it measures semantic rather than merely motion-level differences. The paper reports GGSD roughly one tier above the strongest baseline on Ant and G1 and 28% higher on Franka, averaged over three seeds; the METRA baseline uses a 5-dimensional continuous skill space evaluated at five one-hot basis vectors.

Low-level skills learned in games can be reused directly by humans on tasks outside the training distribution without additional policy training. The training games are sumo pushing, fencing, air hockey and boxing, whereas evaluation uses Maze navigation and CubePush object manipulation, and training happens in Isaac Lab while evaluation happens in MuJoCo web, so both task and simulator shift. Human evaluation with 5 participants (including 2 authors) and 5 trials per task reports 100% success at 82.7 s average for AntMaze, 100% at 62.2 s for AntCube and 84% at 22.0 s for G1Maze; separately, 4 users each played five games against the final 70k checkpoint on AntSumo and humans won only 2 games.

The result targets researchers and interaction designers who want to control high-degree-of-freedom embodied agents with a few discrete inputs, in settings where the skill set is shaped by simple 1v1 competitive games and the downstream task's required behaviors fall within what those games demand. It lets humans compose skills to complete unseen tasks such as Maze and CubePush without training a new controller, and it offers a reusable training recipe for treating game rules as a skill-discovery guidance signal; the authors treat scaling to games that require broader motor capabilities as an important direction toward more general skill repertoires.

The human evaluation involves 5 participants and 5 trials per task, including authors, so success rates and completion times reflect playability in a small sample; semantic consistency of skills is argued only under an assumption of a small number of state regimes, and the trade-off between the number of regimes and interface interpretability remains open; the skill repertoire is bounded by what the game requires, and whether capabilities the game does not demand, such as dexterous manipulation, can emerge is not yet evidenced; the VLM semantic-diversity metric also depends on specific models and prompts, so its stability across models is worth watching.