Player-1: Learning to Play from Synthetic Experience
Player-1 uses synthetic RL environments to substantially improve game-control agents beyond supervised training.
Player-1 is a post-training framework that teaches computer-use models game control from synthetic experience rather than demonstrations alone. It extends the action space with relative pointer motion, sustained inputs, and composite actions, then fine-tunes on 6,044 curated trajectories before PPO training in Player-1-Gym, sixteen synthetic 2D and 3D environments. Player-1-9B and Player-1-35B-A3B improve MineDojo-300 success over their supervised policies by 27.8 and 17.7 percentage points; Player-1-9B reaches 38.33% versus 12.63% for the strongest evaluated baseline, with further gains on ViZDoom-10 and GameWorld.
- Player-1 post-trains computer-use models with synthetic gameplay experience.
- Supervised fine-tuning uses 6,044 curated gameplay trajectories.
- Player-1-Gym has sixteen synthetic 2D and 3D environments trained with PPO.
- Player-1-9B reaches 38.33% MineDojo-300 success versus 12.63% for the strongest baseline.
- Composite keyboard-and-mouse actions increase during reinforcement learning.
- AI models
- Player-1-9BPlayer-1-35B-A3B
From the original article
1 min read at arxiv.orgComputer-use agents have demonstrated strong capabilities in interpreting screens and executing keyboard and mouse actions. However, games, with their visually rich worlds and diverse mechanics, remain challenging and offer a demanding next step toward general digital control.
By Mingyu Ouyang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.