README.md

metadata

pretty_name: WorldOracle Validation Ground Truth

WorldOracle Validation Ground Truth

This public dataset contains the frozen WorldOracle validation split and its ground-truth data. It contains validation episodes only and does not include test-set ground truth.

The data is stored as one uncompressed tar archive per episode. This keeps the repository resumable and avoids hundreds of thousands of individual Hub files. Extract an episode with tar -xf <episode>.tar -C <destination>.

Platforms

Platform	Episodes
AgiBot	20
YAM	15
Human Hand	20
Policy	12
Total	67

validation_manifest.json is the canonical inventory. It records the path, size and SHA-256 of every episode archive. SHA256SUMS covers all repository files except itself. Validation inputs are available under inputs/.

Human Hand absolute pose format

Human Hand validation inputs provide synchronized absolute hand and camera poses in controls/hand_camera_poses.npz. Pose sample t corresponds directly to validation video frame t.

• left_positions_camera and right_positions_camera contain the 21 hand joints in the released camera-reference coordinates.
• left_quaternions_camera_xyzw and right_quaternions_camera_xyzw contain the corresponding joint orientations in XYZW order.
• camera_position_world and camera_quaternion_world_xyzw contain the tracked camera-reference pose.
• left_valid, right_valid, and camera_valid indicate whether each measurement is usable.

The leading initialization interval in which the pose tracking was invalid has been removed, so the first released sample is valid. Any unavailable measurements occurring later in an episode remain present and are marked by the validity arrays. Delta-action tensors, source-frame indices, internal source mappings, and cleaning records are not part of the public release.

Human Hand joint visualization

scripts/visualize_human_hand.py provides a standalone example for projecting the released 3D hand joints onto the rectified top_head frames of a public Human Hand validation episode. It reads the episode's released hand_camera_poses.npz and calibration.json, applies the tracked-camera, physical-fisheye, and virtual-camera transformations, and renders the left- and right-hand 21-joint skeletons next to the original frames.

Install the two runtime dependencies with:

python -m pip install numpy opencv-python

After extracting one of the public Human Hand validation archives, run:

python scripts/visualize_human_hand.py \
  --npz inputs/human_hand/<task>/<episode>/controls/hand_camera_poses.npz \
  --calibration inputs/human_hand/<task>/<episode>/calibration.json \
  --frames-dir <extracted-episode>/video/top_head \
  --output hand_joint_projection.mp4 \
  --fps 15

The script requires the image sequence and NPZ arrays to have identical frame counts and respects the released hand- and camera-validity masks. It contains no hard-coded episode list or camera parameters; all geometry is read from the selected public validation episode.

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}
}