VLA training data from real-world demonstrations
Head- and chest-mounted first-person video of real tasks is exactly the human demonstration data that vision-language-action (VLA) models train on. We collect it in East India, to your protocol.
A vision-language-action model takes in what a robot sees and an instruction in plain language, and outputs the actions to carry it out. To learn that mapping it needs large amounts of demonstration data: real people doing real tasks, filmed from a viewpoint close to the robot's own, with each clip clearly labelled by task.
Lab-recorded demonstrations are clean but narrow. Web video is plentiful but shot from the wrong angle, with no consent trail and no consistent protocol. VLA teams need something in between: real environments, recorded deliberately, at volume, with the paperwork to use it.
What VLA teams need from demonstration data
- The right viewpoint. Egocentric footage shows hands, tools and objects the way an onboard camera will see them.
- Clear task boundaries. Each clip mapped to a task category and step from your protocol, so it can be paired with language instructions.
- Consistency. The same camera settings, mount position and frame rate on every operator and every shift.
- Variety. Different people, sites, objects and lighting, so the model does not overfit to one bench.
- Provenance. Consent and site agreements your legal team can sign off.
How we collect VLA training data
We record to your capture protocol, camera settings and sensor set, so this VLA training data drops straight into your existing data pipeline. In practice that means:
- Protocol review. We go through your task list, instructions and camera settings and flag anything that will be hard to reproduce in the field.
- Site matching. We pick sites in West Bengal where your tasks happen naturally: factories, logistics hubs, garment clusters, kitchens and workshops.
- Operator briefing. Operators are briefed in Bengali, Hindi or English on the protocol and the consent process.
- Capture and checks. Every clip gets frame checks, task tagging and metadata before it passes the approval gate.
- Delivery. Secure upload, in your folder structure and naming convention.
Full rig, metadata and upload details are on our egocentric data collection page.
Task categories that suit VLA training
Task categories are always chosen by you. Examples of what we can capture in East India:
- Fabrication, machining and assembly on working shop floors in Howrah and Liluah.
- Picking, packing, sorting and dispatch in Dankuni and Rajarhat logistics hubs.
- Cutting, stitching and finishing in the Bantala and Metiabruz leather and garment clusters.
- Everyday skilled work in Kolkata kitchens, workshops and garages.
Our sample clips show a packing and dispatch task: wrapping an order, then taping and sealing the carton, recorded head-mounted in 4K. You can watch them on the pilot page.
Pairing demonstrations with language
VLA models learn from demonstrations paired with instructions. Because every clip is tagged with its task and step from your protocol, you can attach your own instruction text or annotation layer downstream. If your protocol calls for spoken narration or on-screen markers during capture, we brief operators to do it consistently.
Compliance built in
Every participant gives written informed consent before a frame is recorded, every site is covered by a written agreement, and every clip carries an audit trail. Workflows are designed with India's DPDP Act in mind. See DPDP Act AI training data compliance.
Related use cases
The same footage often serves more than one model. Many teams use it as imitation learning data for policy training and as humanoid robot data for manipulation research.
Try it on your tasks
The 50–100 hour pilot is the quickest way to see whether our VLA training data fits your pipeline. You pick the tasks, send equipment and guidelines, and we deliver within 2 weeks.