What makes good world-model data
Short clips teach appearance. Long sessions teach consequences: a cloth stays wet after it is rinsed, a stacked box stays on the shelf, a coiled cable keeps its shape. Continuity matters more than polish.
- Uncut sessions long enough to contain full tasks and their after-effects
- Action-rich content where hands change the state of the scene
- Egomotion signals (IMU) aligned with video for action-conditioned prediction
- Diversity of environments, lighting and objects to avoid memorising a single room
How DataClue fits
Our office, food-service and education sessions are among the longest in the catalog, with published samples between 9 and 19 minutes. Sessions captured on GoPro HERO cameras embed accelerometer and gyroscope streams in the same MP4, so egomotion is available without extra hardware.
Volumes are large enough for pretraining: the catalog holds more than 800,000 hours, and per-dataset inventory is labelled as verified or indicative so you can plan around confirmed stock.
Delivery
Video is delivered as original-quality MP4 (H.264 or H.265) with per-session metadata, to your cloud bucket or on physical drives for very large orders.
