Egocentric data collection services for Physical AI
We put head- and chest-mounted cameras on skilled people doing real work in East India, record to your protocol, and ship consented, quality-checked clips with the metadata your pipeline expects.
Egocentric data collection is the work of recording first-person video while a person performs a real task: assembling a part, packing an order, cooking a dish, repairing a machine. The camera sits on the head or chest, so the footage shows hands, tools and objects from the same point of view a robot's onboard camera will have at deployment. That viewpoint is why egocentric video has become core training material for Physical AI, VLA models and imitation learning.
Orchis Green is the field-operations team that makes this happen on the ground in West Bengal. We find the sites, agree access, recruit and brief the operators, run the capture and check every clip before it leaves us.
What we do, end to end
- I
Egocentric video collection
Head- and chest-mounted egocentric data collection of real tasks, recorded to your protocol and camera settings.
- II
Site onboarding & access negotiation
We find suitable sites, agree access with owners and managers, and handle the paperwork.
- III
Field operator recruitment & management
We recruit, brief, schedule and supervise the people who wear the cameras.
- IV
Quality control with audit trails
Frame checks, metadata and secure upload, with the same approval-gate discipline we use in SteelBox.
- V
Hardware deployment support
On-site setup, operation and support for your robot evaluation runs.
Capture specifications
Every project is recorded to your capture protocol. You decide the camera, the mount, the resolution and the frame rate; we make sure it is applied the same way on every operator and every shift. The table below shows what we handle and what you set.
| Item | How it works |
|---|---|
| Rigs | Head-mounted and chest-mounted rigs, worn by trained field operators. Your cameras and sensors, or a setup we agree with you. |
| Resolution | Up to 4K, as in our sample clips. Set by your protocol. |
| Frame rate | Set by your protocol and locked per project so every clip matches. |
| Additional sensors | Whatever your sensor set includes, mounted and synchronised to your guidelines. |
| Clip metadata | Operator ID, site, task, consent reference and capture time on every clip, plus any extra fields your schema needs. |
| Privacy | Face-blurring on request, before delivery. |
| Delivery | Secure upload in your folder structure and naming convention. |
Metadata on every clip
A clip that cannot be traced is a clip your legal and data teams cannot use. Each one carries at least:
- Operator ID: who wore the camera, linked to their signed consent.
- Site: which factory, warehouse, kitchen or workshop, covered by a written site agreement.
- Task: the task category and step you defined in your protocol.
- Consent reference: the consent record that covers the participant and the footage.
The secure upload flow
- Capture. Operators record on site to the agreed protocol, under a supervisor.
- Frame checks. We review each clip for framing, exposure, occlusion and task completeness. Clips that miss the bar are flagged, not shipped.
- Metadata and consent match. Every clip is tagged and matched to its operator, site and consent record.
- Approval gate. A clip only moves to delivery once it passes review. Same rule as SteelBox: AI proposes, humans confirm, every decision is traceable.
- Secure upload. Approved clips are uploaded to your storage with an audit trail.
Where we capture
Our sites span fabrication and assembly in Howrah and Liluah, picking and dispatch in Dankuni and Rajarhat, cutting and stitching in the Bantala and Metiabruz leather and garment clusters, and everyday skilled work in Kolkata kitchens, workshops and garages. These are environments most models have not seen yet, and no major data collector covers the region.
Consent first, always
Every participant gives written informed consent in Bengali, Hindi or English before a single frame is recorded. Workflows are designed with India's Digital Personal Data Protection Act, 2023 in mind, and footage is used only for your model training. Read how it works on our DPDP Act AI training data compliance page.
Start small
The fastest way to test egocentric data collection with us is the 50–100 hour pilot. You choose the task categories, send your equipment and guidelines, and we deliver consented, quality-checked data within 2 weeks.