Guide
What training data do VLA models need? Human demonstrations of real work
VLA models need demonstrations that tie what is seen to what is done, and first-person video of real work is the human layer of that mix: Dovra supplies about 5,000 hours of it per 4-week cycle, with consent and provenance per clip. Robot teleoperation and simulation cover the other two layers. This guide explains what each layer contributes, what human demonstrations add that the others do not, and what to put in a brief so the footage fits your training pipeline.
Short answer
- What a VLA learns from
- Demonstrations linking vision, a task description and actions
- Three data layers
- Human (first-person capture), robot (teleoperation), world (simulation)
- What Dovra covers
- The human layer only — not simulation, not teleoperation
- Standing capacity
- About 5,000 hours per 4-week cycle, 32 countries, deepest in Latin America
- Catalogue
- 15 work domains, 84 task types
- Speed
- 72 hours from brief to first clips; a new domain in 2–3 weeks
What is a VLA model, and why does it depend on demonstrations?
A vision-language-action (VLA) model takes camera input and an instruction and outputs actions, so it learns from demonstrations: recordings of a task being done, paired with what the task is. The model has to connect three things — what the scene looks like, what the goal means and which movements achieve it. Text and still images teach the first two. Only demonstrations teach the third, and the quality and variety of those demonstrations set the ceiling on what the policy can do in an unfamiliar room.
Which data layers go into a VLA training mix?
Most teams combine three layers, and each one answers a different question. Simulation (the world layer) is cheap to scale and good for physics and edge cases, but it inherits the gaps of whoever built the scene. Teleoperation (the robot layer) records the robot's own body and action space, which is precise but slow and expensive per hour. Human first-person capture (the human layer) shows how real people do real tasks in real workplaces, at a volume and variety the other two struggle to reach.
| Layer | Source | Main strength | Main limit |
|---|---|---|---|
| World | Simulation | Scale, control, rare cases | Only as varied as the scenes built |
| Robot | Teleoperation | Exact robot actions | Slow and costly per hour |
| Human | First-person video of real work | Real task order, hands, scene variety | Human body, not the robot's |
What do human demonstrations add that robot and simulation data do not?
Human demonstrations add four things that are hard to get any other way. Hand-object interaction: a head-mounted camera at eye level shows how hands approach, grasp, hold and release tools and materials. Real task order: an unscripted shift shows the steps a job actually takes, including the small preparatory ones. Recovery from mistakes: people drop things, regrip and correct, and a policy needs to see that. Scene variety: cramped kitchens, cluttered workshops and badly lit storerooms that rarely appear in curated datasets.
Why does it matter that the work is real and not staged?
Staged demonstrations look clean and generalize poorly, because they remove exactly the variation a deployed model will meet. Footage of people doing their actual work carries hesitations, workarounds and the messy layouts of real businesses. That is why Dovra records only real, complete tasks inside working businesses, with natural movement, and rejects footage that is purposeless, repetitive or nearly static before it ever counts as an approved hour.
What should you specify when you brief human demonstration data?
A good brief defines five things so the footage matches your training goals. Tasks: the domains and task types you need, ideally from a fixed taxonomy. Scenes: the kinds of workplaces and the country mix. Clip length: short single actions or long multi-step sequences. Camera placement: head-mounted at eye level, tilted down, both hands in view. Exclusions: what must never appear — for example minors, documents, screens, internal pricing and identifiable customers.
What metadata should ship with each clip?
Each clip should carry enough metadata to filter, stratify and audit the dataset. At Dovra that means domain and task from a fixed taxonomy, scene type, country, device model and OS, duration, timestamp, session ID and a human quality rating. Because contributors record on their own phones, the exact device model is logged per clip, so you can stratify by sensor if your pipeline needs it. Annotation, such as step segmentation or your own labels, is available on request.
How do you prove the data is cleared for training?
You prove it with a chain of custody that reaches every clip. Dovra keeps a signed agreement per contributor, written permission per business where filming happened, a payment record per approved hour, a training licence and a provenance document. Delivery is an encrypted transfer to your bucket with a manifest and a dataset card. For the full due-diligence list, see our 10 questions to ask a vendor.
How fast can a human data layer scale?
A standing network scales faster than a one-off collection. Dovra keeps capacity of about 5,000 hours per 4-week cycle across 32 countries, with the deepest bench in Latin America, and goes from brief to first clips in 72 hours in a covered domain. A new domain takes two to three weeks. If you are new to the format, start with what egocentric video data is, and read what drives the price per hour before you budget.
Bottom line
VLA models learn from demonstrations, and the human layer is where task order, hand-object interaction, recovery and scene variety come from. Dovra supplies first-person video of real work at scale: about 5,000 hours per 4-week cycle, with consent and provenance per clip. Dovra is backed and operated by Simera, a remote-talent company for Latin America — read why Simera backs Dovra.
See the human layer before you commit
Tell us the tasks and scenes your VLA needs and roughly how many hours. We reply within one business day with a free sample pack carrying the same metadata as a full delivery.
Request a sample pack See dataset coverageFAQ
What is VLA model training data?
Demonstrations that connect what a camera sees, what the task is called and what the hands or the robot actually do. It usually mixes three layers: human first-person video of real work, robot teleoperation data and simulation.
Can a VLA model learn from human video if the robot has a different body?
Human video does not replace robot data, but it teaches what a policy cannot get cheaply elsewhere: how real tasks are ordered, how hands approach and grasp objects, how people recover from mistakes and how varied real workplaces are. Robot data then grounds that knowledge in the robot's own actions.
Does Dovra provide teleoperation or simulation data?
No. Dovra supplies only the human layer: first-person video of real work, captured head-mounted on the contributor's own phone, with consent and provenance per clip.
How do I get a sample of first-person training data?
Request a free sample pack at dovra.ai. It carries the same metadata and consent record as a full delivery, and Dovra replies within one business day.