WorldOracle/worldoracle-trainset

WorldOracle Training Dataset

This repository provides the training data used for WorldOracle model adaptation on the AgiBot G2 and YAM embodiments.

AgiBot G2

The hosted AgiBot G2 release contains 1,779 complete trajectories at 30 FPS, with synchronized top-head, left-hand, and right-hand videos. Each episode includes camera intrinsics and extrinsics, together with aligned robot actions and states. Private acquisition metadata and shared robot kinematics are intentionally excluded.

YAM: selected ABC-130K episodes

WorldOracle uses a fixed 2,089-episode subset of lerobot/abc_130k_v3_train:

• Fruit manipulation: 1,059 episodes
• Drawer manipulation: 302 episodes
• Towel manipulation: 728 episodes
• Total duration: 2,100 minutes (35 hours)

The authoritative episode list is provided in selected_episode_indices.json. The corresponding public source-shard and timestamp mapping is provided in abc_130k_selected_manifest.json.

Download only the selected episodes

Install the two Python dependencies and make sure ffmpeg and ffprobe are available on PATH:

python3 -m pip install -U huggingface_hub pyarrow

Run the downloader from this repository:

python3 download_abc130k_selected.py \
  --output ./yam_abc130k_worldoracle_selected

The default command exports all 2,089 selected episodes. To verify the pipeline with one episode first:

python3 download_abc130k_selected.py \
  --episode 50 \
  --output ./yam_abc130k_test

Use --workers N to control parallel source-shard downloads. The default is four workers.

The upstream LeRobot v3 dataset stores multiple episodes in shared Parquet and MP4 shards. The script downloads only source shards referenced by the WorldOracle selection, extracts the requested episode rows and timestamp ranges, and removes temporary source shards. The final output contains only the requested WorldOracle episodes.

Each exported episode has the following structure:

episodes/
└── ep000050/
    ├── trajectory.parquet
    └── videos/
        ├── top.mp4
        ├── left_wrist.mp4
        └── right_wrist.mp4

The exporter preserves the original ABC-130K episode index, action, state, frame index, task index, and timestamps. Each extracted video is validated against the episode frame count.

When using the YAM subset, please also follow the terms and citation instructions of the upstream lerobot/abc_130k_v3_train dataset.

License and Citation

All data in this repository are licensed under CC BY-NC-SA 4.0. If you find our benchmark useful in your research, please consider citing our work.

@article{duan2026worldoracle,
  title={WorldOracle: Benchmarking Action-Conditioned World Models against Real-World Executions},
  author={Duan, Haoyi and Chen, Liliang and Wang, Nan and Zhang, Yang and Lu, Yuhang and Chen, Xiaotong and Tseng, Wei-Cheng and Gao, Shenyuan and Li, Zhaoshuo and Jin, Jingyi and others},
  journal={arXiv preprint arXiv:2610.18826},
  year={2026}
}