| --- |
| license: apache-2.0 |
| pretty_name: OmniGameArena |
| task_categories: |
| - reinforcement-learning |
| tags: |
| - games |
| - benchmark |
| - agents |
| - game-playing |
| - unreal-engine |
| viewer: false |
| --- |
| |
| # OmniGameArena |
|
|
| [**Project Page**](https://mxlin043.github.io/OmniGameArena/) | [**Paper**](https://huggingface.co/papers/2606.09826) | [**GitHub**](https://github.com/mxlin043/OmniGameArena) |
|
|
| **OmniGameArena** is a unified benchmark for evaluating game-playing vision-language model (VLM) agents across a diverse suite of 12 interactive video games built with **Unreal Engine 5**. |
|
|
| This repository provides the **packaged, ready-to-run game environments** used by the benchmark. |
|
|
| ## Contents |
|
|
| | File | Platform | Size | |
| |------|----------|------| |
| | `OmniGameArena_Windows.zip` | Windows (x64) | ~2.8 GB | |
| | `OmniGameArena_Linux.zip` | Linux | ~2.9 GB | |
|
|
| Each archive is a single self-contained build of the OmniGameArena environment for the corresponding platform. |
|
|
| ## Games and Modes |
|
|
| OmniGameArena contains **12 games**, spanning three interaction modes: |
|
|
| - **Solo** (7 games) — ObstacleRun2D, ObstacleRun3D, LastStand, MonsterShoot, SceneEscape, CueChase, SoloCraft. Measures an agent's individual ability. |
| - **PvP (Player vs. Player)** (3 games) — SkyDuel, CrystalGuard, MidlineClash. Two agents compete head-to-head. |
| - **Coop (Cooperative)** (2 games) — SharedFloor, HandoffRun. Multiple agents coordinate to succeed together. |
|
|
| ## Sample Usage |
|
|
| To use these environments with the benchmark agents, first download and launch the UE5 environment build, then use the agent code from the [GitHub repository](https://github.com/mxlin043/OmniGameArena). |
|
|
| ### 1. Install the agent code |
| ```bash |
| conda create -n omnigamearena python=3.10 |
| conda activate omnigamearena |
| pip install -r requirements.txt |
| ``` |
|
|
| ### 2. Launch the Environment |
| Launch the downloaded executable for your platform. The environment waits for the agent over TCP (default `127.0.0.1:12345`). |
|
|
| ### 3. Run a benchmark |
| Point the runner at the running environment's host and port: |
| ```bash |
| python scripts/run_benchmark.py \ |
| --config configs/vlm/cold_start/solo/obstacle_run_2d/vanilla_pdq.yaml \ |
| --host 127.0.0.1 --port 12345 |
| ``` |
|
|
| ## Controls |
|
|
| The environment can also be played manually: |
|
|
| | Key | Action | |
| |-----|--------| |
| | `P` | Open the map selector | |
| | `R` | Reset the current game | |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{lin2026omnigamearena, |
| title = {OmniGameArena: A Unified UE5 Benchmark for VLM Game Agents with Improvement Dynamics}, |
| author = {Lin, Mingxian and Qian, Shengju and Liu, Yuqi and Huang, Yi-Hua and Wang, Yiyu and Huang, Wei and Li, Yitang and Zhang, Fan and Hu, Zeyu and Zhu, Lingting and Wang, Xin and Qi, Xiaojuan}, |
| journal = {arXiv preprint arXiv:2606.09826}, |
| year = {2026} |
| } |
| ``` |
|
|
| ## License |
|
|
| This dataset is released under the **Apache License 2.0**. |
