Computer-use agents can already interpret screens and execute keyboard and mouse actions, but games remain a demanding next step because of their visually rich worlds and diverse mechanics. The authors of Player-1 argue that current pipelines rely on large-scale demonstrations with limited coverage of dynamic game states, while scalable interactive training environments are hard to obtain. Their solution is a post-training framework that learns game control beyond demonstrations through synthetic experience.

Player-1 first extends the computer-use action space with relative pointer motion, sustained inputs, and composite actions that combine keyboard and mouse. The model is initialized by fine-tuning on 6,044 curated gameplay trajectories that pair recorded actions with task instructions and visual reasoning. It is then trained with PPO in Player-1-Gym, a suite of sixteen synthetic 2D and 3D environments with programmatically verifiable rewards and randomized observations and control dynamics.

The resulting models, Player-1-9B and Player-1-35B-A3B, improve MineDojo-300 success over their supervised-only policies by 27.8 and 17.7 percentage points, respectively. Player-1-9B reaches 38.33% success versus 12.63% for the strongest evaluated baseline, with further gains over supervised fine-tuning on ViZDoom-10 and GameWorld. Analyses show that broader synthetic environment coverage improves transfer, and that use of composite actions increases during reinforcement learning.

Since this is a single arXiv preprint, the results have not yet been peer-reviewed. The authors' central claim is that learning from synthetic experience substantially improves game control beyond supervised initialization, with gains that generalize beyond the RL training suite.