Humanoid robot data for real-world manipulation
For humanoid robot data, we capture the human side of manipulation: hands-on tasks recorded egocentrically so the viewpoint and hand-object interaction match what a humanoid or robot arm camera will see.
Humanoid robots are built to work in spaces designed for people. To learn that work they need to see it done: how hands approach an object, how grip changes during a task, how tools are picked up and put down, how a person recovers when something slips. That knowledge lives in skilled workers in real workplaces, and the only way to capture it is to record them doing the job.
What we capture
Hands-on tasks such as assembly, packing, cooking, sorting and repair, recorded egocentrically with head- and chest-mounted rigs. Examples from our site network in West Bengal:
- Assembly and fabrication on working shop floors in Howrah and Liluah.
- Packing, sorting and dispatch in Dankuni and Rajarhat logistics hubs, like the wrapping and carton-sealing clips on our pilot page.
- Cutting, stitching and finishing of soft and deformable materials in the Bantala and Metiabruz leather and garment clusters.
- Cooking, cleaning and repair in Kolkata kitchens, workshops and garages.
Task categories are always chosen by you.
Why egocentric video for humanoids
A humanoid's cameras usually sit at or near head height, looking down at its hands. Head-mounted footage reproduces that view almost directly. Chest-mounted rigs give a steadier view of the workspace in front of the body. Many teams use both. Either way, the footage shows hand-object interaction from the angle the robot will work from, which third-person video cannot do.
Skilled hands, varied environments
Manipulation data is only as good as the hands in it. We record people who do these tasks for a living: fabricators, packers, tailors, cooks and mechanics. Their movements are fluent and efficient, and they handle the small recoveries and adjustments that a scripted demonstrator in a lab never shows. Because our sites span factories, warehouses, garment clusters and small kitchens, the same task can be captured with different tools, objects, surfaces and lighting.
That variety matters for humanoids in particular. A robot that will work in homes, shops and factories needs to have seen more than one bench. Field operations in East India cost less to run than in the crowded metros, so you can afford more of that variety for the same budget.
Robot evaluation support
Collecting data is half the job. Once you have a policy, you need to test it somewhere real. We also support on-site hardware deployment for your own robot evaluation runs: on-site setup, operation and support at the same kinds of sites where the training data was captured. That means you can evaluate your humanoid or robot arm in the environments it learned from.
How a humanoid data project runs
- Scope. You share the manipulation tasks, capture protocol, camera settings and sensor set.
- Sites and operators. We select suitable sites, agree access and recruit operators who do the work for a living.
- Capture. Operators record to your protocol under supervision, with consent signed beforehand.
- QC and approval. Frame checks, metadata and an approval gate on every clip.
- Delivery. Secure upload with operator ID, site, task and consent reference per clip.
See full capture specifications on the egocentric data collection page.
Related training approaches
Humanoid robot data is often used as imitation learning data for policy training, or paired with instructions as VLA training data. We record to your protocol, so the same capture can serve both.
Compliance
Every participant gives written informed consent in Bengali, Hindi or English, every site has a written agreement, and every clip carries an audit trail. Read our approach to DPDP Act AI training data compliance.
Start with a pilot
The 50–100 hour pilot lets you test our humanoid robot data on your own manipulation tasks before committing to a larger programme. You pick the tasks, send equipment and guidelines, and we deliver within 2 weeks.