EmbodiedCity/iWorld-Bench-Dataset

iWorld-Bench Simulation Archives

News: Congratulations to the iWorldBench team! iWorldBench has been accepted to ICML 2026.

Important Dates

• Paper accepted: May 1, 2026
• Code release: May 18, 2026
• Dataset release: May 19, 2026

iWorldBench is a benchmark for evaluating camera-controllable video generation models and interactive world models. This dataset repository hosts the packaged simulation archives associated with iWorld-Bench. It provides rendered simulation videos and aligned camera-trajectory files organized by simulator environment, complementing the evaluation code, packaged metadata, trajectory resources, and reference inference adapters released in the main repository.

The uploaded simulation data is organized by environment index from dataset_env0 to dataset_env26. Each environment is distributed as a collection of independent ZIP archives together with a manifest file, which makes large-scale upload, download, partial retrieval, and recovery more robust.

Overview

The simulation archives follow the simulator setting described in the iWorld-Bench paper. The benchmark selects high-quality environments from four simulator sources:

• Aerial VLN
• UAV ON
• Openfly
• EmbodiedCity

In this release, dataset_env0 corresponds to the EmbodiedCity environment. The remaining dataset_env* folders correspond to selected simulator environments within the uploaded 0-26 environment index range.

This data is intended for research on interactive world models, camera-controlled video generation, trajectory following, action-conditioned visual prediction, and embodied world simulation.

Dataset Contents

Each extracted environment contains two aligned components:

Component	Description
videos/	Rendered simulation video clips collected from selected simulator environments.
cameras/	Camera-trajectory files aligned with the rendered videos by filename stem.

The videos/ directory provides visual observations, while the corresponding cameras/ directory provides camera-control trajectories that can be used by models that consume intrinsic/extrinsic camera parameters, trajectory commands, or related camera-control signals.

A typical paired sample uses the same filename stem:

videos/example_name.mp4
cameras/example_name.txt

Directory Structure

The simulation archives are stored under:

sim/archives_split/
├── dataset_env0/
├── dataset_env1/
├── dataset_env2/
├── ...
└── dataset_env26/

Each dataset_env* directory contains several ZIP archives and one manifest file:

dataset_env0/
├── dataset_env0_part000001.zip
├── dataset_env0_part000002.zip
├── dataset_env0_part000003.zip
├── ...
└── manifest.json

After extraction, each environment follows the standard paired-data layout:

dataset_env*/
├── videos/
│   ├── *.mp4
│   └── ...
└── cameras/
    ├── *.txt
    └── ...

Archive Format

The dataset is stored as split-by-environment ZIP archives instead of one monolithic archive. This layout is designed for practical large-scale distribution: users can download only the environments they need, resume or retry individual archives more easily, and verify each environment independently.

Example archive names:

dataset_env0_part000001.zip
dataset_env0_part000002.zip
dataset_env0_part000003.zip
...

Each ZIP file is an independent archive. The accompanying manifest.json records archive-level metadata such as archive names, file counts, original paths, and file sizes. It can be used to check whether all expected files have been downloaded and extracted.

Environment Naming

The released simulation subset uses the following environment naming convention:

dataset_env{ID}

where {ID} is the environment index in the uploaded range.

Confirmed mapping:

dataset_env0 = EmbodiedCity

Other dataset_env* folders correspond to selected high-quality environments from the simulator pool used by iWorld-Bench.

Example Extraction

To extract one environment:

ENV_DIR=sim/archives_split/dataset_env0
OUT_DIR=extracted_env0

mkdir -p "$OUT_DIR"

for zip_file in "$ENV_DIR"/*.zip; do
    unzip -n "$zip_file" -d "$OUT_DIR"
done

After extraction, the expected structure is:

extracted_env0/
└── dataset_env0/
    ├── videos/
    └── cameras/

Suggested Usage

The recommended usage is to read paired video and camera files by matching their filename stems. The video clip provides the visual sequence, and the corresponding camera file provides the aligned camera-control trajectory.

This simulation subset can be used for:

• Training interactive world models: use paired videos and camera trajectories as simulation supervision.
• Evaluating camera-controlled video generation: test whether generated videos follow a specified camera path.
• Studying trajectory-following ability: analyze the consistency between visual motion and target camera trajectories.
• Building embodied world-model benchmarks: construct evaluation protocols for action-conditioned and camera-conditioned world simulation.

For the official iWorldBench evaluation workflow, please refer to the main code repository and metric documentation:

• GitHub Code
• Leaderboard
• Metric documentation

Inference Dataset (dataset/all_pack)

The dataset/all_pack directory contains the packaged inference data for running iWorld-Bench evaluations. It includes:

File / Directory	Description
assets/	First-frame images organized by source dataset and task (Diff/Mem).
metadata.csv	Main metadata for Diff and Memory tasks, with sample IDs, dataset names, file paths, control types, and difficulty levels.
camera_following_metadata.csv	Metadata for the Camera-Following task.
videoxfun_diff.csv	VideoX-Fun-compatible Diff inference CSV.

This dataset is designed for inference: use it together with the evaluation code in the main repository to generate videos with your model and then run the iWorld-Bench evaluation pipeline.

To download and use this dataset for inference:

# Install huggingface_hub if needed
pip install huggingface_hub

# Download the dataset
huggingface-cli download EmbodiedCity/iWorld-Bench-Dataset --repo-type dataset --local-dir iWorld-Bench-Dataset

Citation

If you use this dataset, please cite the iWorld-Bench paper:

@misc{fang2026iworldbenchbenchmarkinteractiveworld,
      title={iWorld-Bench: A Benchmark for Interactive World Models with a Unified Action Generation Framework}, 
      author={Jianjie Fang and Yingshan Lei and Qin Wan and Ziyou Wang and Yuchao Huang and Yongyan Xu and Baining Zhao and Weichen Zhang and Chen Gao and Xinlei Chen and Yong Li},
      year={2026},
      eprint={2605.03941},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.03941}, 
}