Player-1 triples game-agent success via synthetic-env RL

Player-1-9B reaches 38.33% success on MineDojo-300, versus 12.63% for the strongest evaluated baseline. That is roughly a threefold jump, and it comes not from more human gameplay data but from reinforcement learning inside sixteen synthetic environments the authors built from scratch.

The preprint on arXiv positions games as a stress test for general computer-use agents. "Computer-use agents have demonstrated strong capabilities in interpreting screens and executing keyboard and mouse actions," the abstract reads. "However, games, with their visually rich worlds and diverse mechanics, remain challenging and offer a demanding next step toward general digital control."

Their recipe is two-stage. Supervised fine-tuning runs on 6,044 curated gameplay trajectories pairing recorded actions with task instructions and visual reasoning. Then PPO reinforcement learning runs across Player-1-Gym, described in the paper as "sixteen synthetic 2D and 3D environments designed around control demands shared across games," with "programmatically verifiable rewards" and randomized observations. The action space itself is extended with "relative pointer motion, sustained inputs, and composite actions."

Over the matching SFT policies, Player-1-9B gains 27.8 percentage points on MineDojo-300 and Player-1-35B-A3B gains 17.7. The paper reports further gains over SFT on ViZDoom-10 and GameWorld, and flags a specific pattern during training: "broader synthetic environment coverage improves transfer, while the use of composite actions that combine multiple keyboard and mouse inputs increases during RL."

The abstract names no compute cost, no commercial-game evaluation, and no comparison against RL-from-demo systems from other labs. All reported transfer stays within the paper's own benchmark suite.