Player-1 is an open VLM post-training framework for generalist game-playing agents, following:
SFT on Player-1-Data → RL in Player-1-Gym → Evaluation on held-out gameplay benchmarks.

---

Core components of the Player-1 framework:

• Player-1-Data: 6,044 curated trajectories from 172 games, complete with screenshots, chain-of-thought reasoning, and normalized keyboard/mouse actions.
• Player-1-Gym: 16 synthetic 2D/3D environments for aiming, combat, driving, navigation, and UI crafting with verifiable rewards.
• Training Engine: VeRL integration supporting PPO, GRPO, and GSPO. Base models: Qwen3.5-9B / Qwen3.5-35B-A3B.
• Gameplay Evaluation Suite: Evaluation across Player-1-Gym, MineDojo (300 tasks), ViZDoom (10 maps), and GameWorld (170 tasks).

🎮 RL Training Environment Previews

2D Aiming	2D Shooting	2D Collect & Dodge	2D Lane Driving
2D Circuit Driving	2D Platforming	2D Crafting	3D Aiming
3D Shooting	3D Road Driving	3D Lane Driving	3D Spear Duel
3D Sword Duel	3D Melee Combat	3D Navigation	3D Chasing

player1-introduction.mp4

---

📢 News

• [2026.11] Released Player-1 models, Player-1-Data, and the complete Player-1 codebase: Player-1-Gym, training recipes (SFT & RL), and benchmark evaluators.
• [2026.10] Released the paper (arXiv 2610.20584) and project page.

---

⚡ Quick Start

Play the 16 Gym environments directly in your browser. Use Python 3.11+ and run from the repository root:

python -m pip install -r requirements/gym.txt
python -m playwright install chromium
PYTHONPATH=. python examples/gym/play.py

Open the printed URL in your browser to play manually.

For Python programmatic stepping, run examples/gym/step.py:

import asyncio
from player1.gym import GymEnv

async def main():
    async with GymEnv("base_3d_aim") as env:
        obs = await env.reset(seed=42)
        # Relative mouse movement [-1000, 1000] and click
        obs = await env.step({
            "type": "game_step_action", "id": "control", "duration_ms": 200,
            "mouse": {"mode": "relative", "dx": 50, "dy": -20},
            "mouse_click": {"button": "left"},
        })
        print(f"Reward: {obs.reward}, Terminated: {obs.terminated}")

asyncio.run(main())

---

📚 Guides & Documentation

Workflow	Guide / Entry Points	Description
Gym	Setup and Python API	16 environments, continuous initialization, and action specs
Data	Reading and format · Production and Viewer	Parquet schema, multi-turn history reader, and VLM annotation
Inference	Model server and examples	vLLM serving, prompt templates, and single-frame/interactive agent
Training	SFT → RL guide · SFT launcher · RL launcher	veRL integration, PPO/GRPO/GSPO, and checkpoint export
Evaluation	Benchmark setup and trajectory viewer	MineDojo-300, ViZDoom-10, GameWorld, and web-based viewer
Sources	Component sources and licenses	Upstream projects and asset notices

---

Player-1 Overview

Player-1 training and evaluation pipeline: Computer-use model → SFT on Player-1-Data → RL in Player-1-Gym → Evaluation in external games

1/ Player-1-Data

We curate 6,044 gameplay trajectories from 172 games, with recorded keyboard and mouse actions mapped to a shared game-control action space. Each training example includes task instructions, screenshot observations, reasoning, and actions for supervised learning.

2/ Player-1-Gym

Sixteen synthetic 2D and 3D environments cover aiming, combat, driving, navigation, and UI interaction. All use screenshot observations and keyboard and mouse controls, with verifiable rewards computed directly from game state. Randomizing appearance, camera, scene, and control broadens the agent's training experience.

3/ Evaluation Benchmarks

We evaluate transfer to games outside the RL training suite:

Minecraft	ViZDoom	GameWorld

• MineDojo (MineDojo-300, 300 tasks, 1000 turns per step): Goal-directed control in Minecraft's open-ended world.
• ViZDoom (ViZDoom-10, 10 maps, 150–2,100 turns per step): First-person navigation, aiming, and combat in ViZDoom.
• GameWorld (170 tasks, 400 turns per step): Game control across 34 browser games with diverse mechanics.

---

🚀 Release Status

• Launch: Paper & Project Page
• Demonstration SFT Dataset: Hugging Face.
• Model Checkpoints: Player-1-9B and Player-1-35B-A3B.
• Player-1-Gym: 16 synthetic 2D/3D environments with Python Playwright bindings.
• Player-1-Data Pipeline: Ingestion, action standardization, and VLM annotation.
• Training Framework: VeRL-based SFT and RL (PPO / GRPO / GSPO) recipes for Qwen3.5-9B / Qwen3.5-35B-A3B.
• Evaluation & Benchmark: Player-1-Gym, MineDojo-300, ViZDoom-10, and GameWorld.

---

📑 Citation

If you find our work or dataset useful, please consider citing:

@misc{ouyang2026player1learningplaysynthetic,
  title={Player-1: Learning to Play from Synthetic Experience},
  author={Mingyu Ouyang and Yiran Wang and Siyuan Hu and Xingyilang Yin and Huiqiang Sun and Guangzhi Wang and Junfu Pu and Cheng Cheng and Honghao Chen and Ying Shan and Mike Zheng Shou and Hwee Tou Ng},
  year={2026},
  eprint={2610.20584},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.20584}
}