Kuration SDK
A Python SDK for curating world-model training data. Turns raw gameplay, robot, and egocentric footage into reproducible dataset shards. Shipped operators handle action-state consistency, collision oversampling, and SLAM trajectory recovery.
What it does. Kuration turns large Physical AI datasets, time-series logs or video, into training-ready data for world models. Point it at the raw footage and it picks the windows worth training on, writes them as shards, and drops a config alongside so the same run can be rebuilt.
How it works. The pipeline is built from operators: named steps that measure a per-window feature, filter windows out, or select which windows to keep. Chain them in Python, call explain() to see what the pipeline will do, then output shards with SHA-256 checksums in the manifest.
Shipped operators for gameplay.
- Action-state consistency: measures the real screen motion behind every key or mouse press and flags clips where the action doesn't reliably cause the state change.
- Collision detection: finds and classifies wall-hits so the rare interaction moments a world model needs can be oversampled.
- SLAM trajectory recovery: reconstructs the camera's 3D path from raw video and scores each window for gaps, impossible motion, and diversity of movement.