Computer-use agents have become adept at reading screens and issuing keyboard and mouse commands, but games remain a difficult frontier because of their rich visuals and varied mechanics. A new research paper, Player-1, proposes a post-training framework that learns game control from synthetic experience rather than relying solely on large-scale demonstrations, which often cover only a narrow slice of dynamic game states.
The authors extend the computer-use action space with relative pointer motion, sustained inputs, and composite actions. For supervised initialization, they fine-tune computer-use models on 6,044 curated gameplay trajectories that pair recorded actions with task instructions and visual reasoning.
They also build Player-1-Gym, a suite of sixteen synthetic 2D and 3D environments designed around control demands shared across games. Programmatically verifiable rewards, randomized observations, and randomized control dynamics enable multi-environment reinforcement learning with PPO.
The resulting models, Player-1-9B and Player-1-35B-A3B, improve MineDojo-300 success over their SFT 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 SFT on ViZDoom-10 and GameWorld.
Analyses indicate that broader synthetic environment coverage improves transfer, and that the use of composite actions combining multiple keyboard and mouse inputs increases during RL. The authors conclude that learning from synthetic experience substantially improves game control beyond supervised initialization, with gains that generalize beyond the RL training suite.