Player-1

Player-1: Learning to Play from Synthetic Experience Mingyu Ouyang, Yiran Wang, Siyuan Hu, Xingyilang Yin, Huiqiang Sun, Guangzhi Wang, Junfu Pu, Cheng Cheng, Honghao Chen, Ying Shan, … arxiv.org [𝚌𝚜.𝙲𝚅 𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙷𝙲] 💬Code: TencentARC/Player-1 t.co

Paper: Player-1: Learning to Play from Synthetic Experience arxiv.org Code: TencentARC/Player-1 Project: player-1-project.github.io

(3/3) See Player-1 in action, or try the synthetic worlds yourself 👇 (more gameplay demos will be release soon!) 🌐 Project: player-1-project.github.io 📄 Paper: huggingface.co 💻 Code: TencentARC/Player-1

Code, models, and datasets coming soon on Hugging Face 🤗 Paper: paperswithcode.co Project: player-1-project.github.io GitHub: TencentARC/Player-1

Introducing Player-1: empowering general VLMs and computer-use agents to master complex gameplay by learning from synthetic experience. @OuyyyangMingyu Hugging Face Daily: huggingface.co GitHub: TencentARC/Player-1 Project Page: player-1-project.github.io · Challenge: Games have become a vital proving ground for AI models to reason, plan, and operate. However, there is no mature open-source framework to boost gameplay capabilities of general VLMs or computer-use agents. Moreover, interactive Game RL remains nearly uncharted. . What: The first framework to unlock Game RL via synthetic environments, releasing a full suite of Player-1 model family across 2B, 4B, 9B, and 35B model sizes. ·How: Bridging a unified keyboard-mouse game action space with supervised warm-up and PPO training across 16 controllable 2D/3D synthetic environments with randomized visuals, physics, and verifiable rewards. . Result: Substantially surpasses existing baselines in task success and out-of-distribution generalization. Without game-specific training, Player-1 generalizes directly to diverse unseen games and AAA titles, e.g., Watch Dogs, Forza Horizon 6, and GTA V (see more demos later!).

Player-1: Learning to Play from Synthetic Experience Computer-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. Current game-agent learning pipelines often rely on large-scale demonstrations with limited coverage of dynamic game states, while scalable environments for interactive training remain difficult to obtain. We introduce Player-1, a post-training framework for learning game control beyond demonstrations through synthetic experience. We extend the computer-use action space with relative pointer motion, sustained inputs, and composite actions. For supervised initialization, we fine-tune computer-use models on 6,044 curated gameplay trajectories that pair recorded actions with task instructions and visual reasoning. We construct Player-1-Gym, sixteen synthetic 2D and 3D environments designed around control demands shared across games. Programmatically verifiable rewards and randomized observations and control dynamics support multi-environment reinforcement learning with PPO. The resulting 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 compared with 12.63% for the strongest evaluated baseline, with further gains over SFT on ViZDoom-10 and GameWorld. Further analyses show that broader synthetic environment coverage improves transfer, while the use of composite actions that combine multiple keyboard and mouse inputs increases during RL. We demonstrate that learning from synthetic experience substantially improves game control beyond supervised initialization, with gains that generalize beyond the RL training suite. Comments: Project Page: player-1-project.github.io and code: TencentARC/Player-1