Egocentric data providers for robotics.
Eight suppliers, compared on the five things that decide a purchase. MOVAS is one of them; the figures are each provider’s own published claims.
| Provider | Supply model | Published scale | Annotation depth | Rights structure | Delivery | Best for |
|---|---|---|---|---|---|---|
| MOVAS AI | Employer-signed third-party workplaces | 100,000 h in inventory | L1–L5 + 4D + MoCap + 3D + retargeting | Employer agreement, then individual consent | UMI · LeRobot · RDT · OXE · native | Contact-rich work with auditable workplace rights |
| Build AI | Instrumented factory workers, open release | 1,000,000 h free (Apache 2.0) | Raw video, no IMU or labels | Open licence | Hugging Face streaming | Free industrial volume for pre-training |
| Awign | India gig workforce | 1,000+ h/day, 1,000+ cities | Detection, segmentation, action labels | Contributor-level | Custom delivery | High volume at the lowest unit cost |
| Lightwheel | Global capture + simulation | 300,000+ h delivered | Video, audio, depth, structured labels | Not published | Sim-ready, robot-agnostic | Breadth across 10,000+ task types |
| Luel | Rights-cleared marketplace | 3,670 h Ego4D subsets · 3M contributors | Audio, gaze, 3D labels | Contributor consent + chain of title | Marketplace download | Fast start on benchmark-derived data |
| Maxinsights | Global contributor network | 2M h accumulated · 450K h/month | Nine layers incl. tactile and force | Not published | Not published | Deepest annotation stack at volume |
| Mecka | Body-sensor capture network | Not published — reported ~$100M annualised | Human motion capture for world models | Not published | Not published | World-model programmes needing human motion at volume |
| EgoData | India partner networks | Not published | Task labels, step segmentation, object tags | Consent-first, participant compensation | Cloud bucket / encrypted transfer | Indian homes, factories and construction sites |
| Appen | Enterprise contributor network | 1M+ contributors | Boxes, segmentation, action labels | Enterprise governance | Enterprise pipelines | Compliance-heavy enterprise programmes |
Figures are as published by each provider on its own site or in named press coverage, current at the time of writing. “Not published” means the provider does not state it publicly — not that it is absent. We include ourselves and have not scored anyone; buyers should verify against their own requirements.
How to read this table
Supply model decides rights
A contributor network licenses what individuals can license. A workplace network licenses what an employer can license. These are different legal objects, and only one of them covers premises, processes and the colleagues in frame.
Inventory is not capacity
Hours already captured ship in days. Monthly capacity means the hours do not exist yet and your schedule depends on someone else’s collection calendar.
Depth is where scarcity moved
Raw hours are no longer scarce — over a million are free. Per-joint contact, metric geometry and physics-validated retargeting still are.
Format is a hidden cost
A cloud bucket of MP4s is not a training set. Native UMI, LeRobot, RDT or Open-X-Embodiment output removes weeks of pipeline work.
Common questions
Egocentric data is first-person video recorded from the viewpoint of the agent performing a task, usually with a head-mounted or chest-mounted camera. It preserves hand position, contact points and gaze — signals a third-person camera cannot recover — which is why it transfers to robot manipulation policies where external footage does not.
They are good enough for pre-training a visual backbone and nothing more. Openly released egocentric corpora now exceed one million hours, but they ship as low-resolution raw video with no inertial stream, no hand pose and no contact labels. Turning that into a trainable trajectory is the work that is still paid for.
Ask who signed. An individual can license their own likeness, but not their employer’s premises, processes, or the colleagues who appear in frame, and most employment agreements restrict on-site recording in the first place. A provider whose rights chain stops at the contributor cannot grant what a workplace dataset actually requires.
Separate hours in inventory from collection capacity, and separate human video hours from robot trajectories. A vendor quoting monthly capacity has not yet captured those hours; a vendor quoting robot trajectories is solving a different problem from one quoting egocentric human video.
Past a few thousand hours, yes. Volume is commoditising — free corpora and low-cost gig networks have removed scarcity from raw hours. What remains scarce is per-joint contact state, metric-scaled geometry, retargeted trajectories validated in simulation, and a rights chain that survives legal review.