The dataset
Vallum Labs collects consent-cleared, action-labeled egocentric video of outdoor manual work and licenses it to robotics and world-model teams. This page states the schema, the consent and provenance stack, the capture specs, and the offer exactly as they stand today.
Annotation schema
- Per-clip task label
- Environment and scene tag
- Geography tag
- Hand-visibility flag
- 21-keypoint hand-pose track where recoverable
- Object-in-hand label
- Action-segment start and stop timestamps
- Consent and provenance ID linking the clip to a signed release
Compatible with the Ego4D and Ego-Exo4D annotation conventions.
Consent and provenance
- Written participant release in English, Afrikaans, and isiXhosa
- Site-owner authorization
- Face blur applied before export. Non-negotiable
- Per-clip provenance record: who, where, when, device, consent ID
- POPIA compliant, including Section 57 on biometric data export
- C2PA content-provenance signing where the pipeline supports it
Capture specs
The capture kit today is four head and chest mounted iPhone rigs in the field, carried in personally. Capture is monocular RGB at 1920x1080, 30 fps, HEVC at capture, with camera settings locked for the whole shift.
Collection window
The first collection window is the Western Cape harvest season, September 2026 to January 2027, agriculture first. Construction, rooftop solar, offshore rigs, wind turbines, and high-rise ironwork are the stated direction beyond agriculture.
The offer
A paid pilot delivers curated, annotated, consent-cleared hours of footage. Licensing defaults to non-exclusive. Time-boxed, task-boxed exclusivity is available as a paid premium, scoped to one task and one window.