WorldOracle/worldoracle-testset

WorldOracle Test Inputs

This public dataset contains model inputs for the frozen WorldOracle test split. It does not contain ground-truth target videos or evaluation results.

Platforms

Platform	Episodes
AgiBot	80
YAM	60
Human Hand	80
Policy	48
Total	268

Human Hand input format

Human Hand test episodes provide frame-0 camera observations, calibration, and synchronized absolute per-frame hand/camera poses. Pose sample t corresponds to episode frame t. Only absolute per-frame poses are provided. Delta-action tensors are not included; consumers should derive any action or transition representation required by their model from the released poses. The left_valid, right_valid, and camera_valid arrays indicate whether each measurement is usable. Some hand pose measurements may be unavailable, for example due to occlusion, and are identified by left_valid or right_valid; camera pose availability is indicated separately by camera_valid. Each episode begins with valid measurements. Internal source mappings, target videos, masks, prompts, and evaluation outputs are not included.

test_manifest.json is the canonical inventory. SHA256SUMS records the checksum of every uploaded file.

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}
}