README.md

metadata

license: apache-2.0
size_categories:
  - 100K<n<1M
task_categories:
  - image-to-video
tags:
  - robotics
  - world-models

MultiWorld Dataset

Dataset Summary

MultiWorld is a large-scale multi-agent multi-view video dataset collected for training video world models. It contains two complementary sources of data:

1. It Takes Two Gameplay Dataset: 100+ hours of real human gameplay from the cooperative action-adventure game It Takes Two, featuring dual-agent synchronized actions with distinct first-person viewpoints.
2. RoboFactory Manipulation Dataset: Multi-robot manipulation trajectories spanning 4 tasks with 2-4 agents and variable camera viewpoints, including both success and failure episodes.

This dataset is the official release accompanying the paper MultiWorld: Scalable Multi-Agent Multi-View Video World Models.

• Homepage: https://multi-world.github.io
• Repository: https://github.com/CIntellifusion/MultiWorld
• Paper: arXiv:2604.18564

---

Sample Usage

Dataset Download

You can download the dataset using the Hugging Face CLI:

hf auth login
hf download Haoyuwu/MultiWorldData --repo-type dataset \
    --local-dir ./data
bash preprocess/untar_chunks.sh

After running preprocess/untar_chunks.sh, the archives are extracted to:

• data/ittakestwo_release/ — It Takes Two dataset
• data/robots_release/ — Robotics dataset

---

Dataset Details

It Takes Two Gameplay

Property	Value
Total Duration	100+ hours
Frame Rate	60 FPS
Resolution	480 × 960
Agents	2 players
Viewpoints	2 distinct first-person views per episode
Actions	Synchronized keyboard and mouse actions per agent
Modality	RGB video + discrete/continuous action vectors

The gameplay videos are captured from real human players cooperating in the game. Each frame is accompanied by per-agent action labels capturing keyboard presses and mouse movements.

RoboFactory Manipulation

Property	Value
Tasks	4 multi-robot manipulation tasks
Agents	2–4 robots per task
Viewpoints	Variable camera configurations per task
Resolution	256 × 320
Success Episodes	1,000 per task
Failure Episodes	2,000 per task
Modality	RGB video + robot proprioception + actions

Tasks include collaborative stacking, pushing, and pick-and-place scenarios. Both successful and failed trajectories are included to support learning robust world models and failure prediction.

---

Possible Usage

The dataset is intended for research in:

• Video world models
• Multi-agent video generation
• Multi-view consistent video generation.

---

Contact

For questions about the dataset, please open an issue on the GitHub repository or contact the authors.

Citation

@article{wu2025multiworld,
  title={MultiWorld: Scalable Multi-Agent Multi-View Video World Models},
  author={Wu, Haoyu and Yu, Jiwen and Zou, Yingtian and Liu, Xihui},
  journal={arXiv preprint arXiv:2604.18564},
  year={2026}
}