Scaling Gameplay Experience for Visual Agents through Environment–Player Co-Development
🌐 Project Website · 💻 GitHub · 🤗 Collection
Game2Player explores how visual agents learn to play games through environment–player co-development. We bring together game environments, gameplay experience, and trained players to study visual decision-making and interactive control.
Our models take game screenshots, task instructions, and recent actions as input, then predict the next keyboard or mouse action.
Explore our visual game-playing models, with weights and usage instructions available in each repository:
We evaluate visual agents on PlayerBench, featuring 50 held-out games and 287 tasks, and GameWorld, featuring 34 games and 170 tasks.
Explore PlayerBench · View the Leaderboard
Visit our GitHub repository for code and benchmark instructions, or open a model card to run a Game2Player model.
Questions, feedback, and contributions are welcome through GitHub Issues.