Physical AI data collection from East India
Physical AI learns from the real world. We capture it: first-person video and sensor data of skilled people doing real tasks in factories, warehouses, kitchens and workshops across West Bengal.
Language models learned from the internet. Robots and embodied AI cannot. A model that has to pick, pack, assemble or cook needs to see those actions done by human hands, in real places, with real clutter and real variation. That data does not exist on the web. Someone has to go out and record it, with consent, to a consistent protocol, at volume.
That is what Physical AI data collection means for us: field operations. We are a team on the ground in East India that turns your capture protocol into hours of usable, consented, quality-checked footage.
Why the data has to be egocentric
Most Physical AI training pipelines want footage that looks like what the robot will see. A head- or chest-mounted camera captures hand-object interaction, gaze and the actor's point of view, which is close to a robot's onboard camera at deployment. Our egocentric data collection service covers the rigs, metadata and upload flow in detail.
Three ways teams use our data
VLA training data
Vision-language-action models map what they see and what they are told to the actions they take. Head- and chest-mounted first-person video of real tasks is exactly the human demonstration data these models train on. We record to your protocol, camera settings and sensor set so clips drop into your existing pipeline. Read more about VLA training data.
Imitation learning data
Imitation learning teaches a policy by example. It needs many clean, consistent demonstrations of the same task by different people in different places. We recruit and brief operators so demonstrations follow your protocol while keeping the natural variation that helps models generalise. Read more about imitation learning data.
Humanoid robot data
Humanoid and robot-arm teams need the human side of manipulation: assembly, packing, cooking, sorting and repair, recorded so the viewpoint matches what the robot's camera will see. We also support on-site hardware deployment for your own robot evaluation runs. Read more about humanoid robot data.
Why East India
Everyone is in Delhi and Bengaluru. We're in the East. That matters for three reasons:
- Lower cost. Field operations in Eastern India cost less to run than in the crowded metros.
- Fresh task environments. The factories, warehouses, kitchens and workshops here are environments most models have not seen.
- Zero competition. No major data collector works in the region yet, so sites and operators are not already committed elsewhere.
See the site types and task examples for Howrah and Kolkata.
What you get
- Footage recorded to your capture protocol, camera settings and sensor set.
- Operator ID, site, task and consent reference on every clip.
- Frame checks and an approval gate before anything ships.
- Written informed consent from every participant and a written agreement for every site.
- Footage used only for your model training and never resold.
Our compliance process is set out on the DPDP Act AI training data compliance page.
Who we work with
We work with robotics AI labs in the US, EU and worldwide that need data from environments no one else has captured. Every engagement is scoped as a written proposal with fees, hours and deliverables agreed before work starts.
How to start
Start with the 50–100 hour pilot in Kolkata and Howrah. You pick the task categories, send equipment and guidelines, and we deliver within 2 weeks. It is also the fastest way to get a firm quote for a larger programme.