Imitation learning data: human demonstrations at volume
Imitation learning pipelines train on human demonstration data. We record those demonstrations egocentrically, in real workplaces across East India, to your protocol and with full consent.
Imitation learning teaches a robot policy by example. Instead of hand-writing rules or rewards, you show the model many examples of a task being done well and it learns to reproduce the behaviour. The quality of the policy depends heavily on the quality of those examples: how consistent they are, how varied, and how closely the viewpoint matches what the robot will see.
What makes good imitation learning data
- Consistency. The same task performed to the same protocol, so the model learns the task rather than the noise.
- Natural variation. Different people, hands, objects, sites and lighting, so the policy generalises beyond one demonstrator.
- Complete episodes. Each demonstration runs from a clear start to a clear end, with failed or partial attempts flagged.
- A matching viewpoint. Head- or chest-mounted footage that sees hands and objects the way the robot's camera will.
- Traceability. Operator, site, task and consent recorded for every episode.
How we collect it
Head- and chest-mounted first-person video of real tasks is exactly the human demonstration data that imitation learning pipelines train on. We record to your capture protocol, camera settings and sensor set, so clips drop into your existing pipeline.
- Define the episode. We agree with you where each demonstration starts and ends, and what counts as a success.
- Recruit for variety. We recruit and schedule operators across sites so you get many demonstrators, not one.
- Brief for consistency. Operators are briefed in Bengali, Hindi or English and supervised on site so the protocol is followed every time.
- Check every episode. Frame checks, completeness checks and metadata before the approval gate.
- Deliver securely. Approved clips are uploaded to your storage with an audit trail.
Details of rigs, metadata fields and upload are on the egocentric data collection page.
Real tasks, real environments
Policies trained only on lab benches struggle in real workplaces. Our sites give you the clutter and variation of real work:
- Fabrication, machining and assembly in Howrah and Liluah factories.
- Picking, packing, sorting and dispatch in Dankuni and Rajarhat logistics hubs.
- Cutting, stitching and finishing in the Bantala and Metiabruz leather and garment clusters.
- Cooking, repair and workshop tasks in Kolkata.
Why collect imitation learning data in East India
Diversity of demonstrators and environments is one of the biggest levers on how well an imitation policy generalises. Most demonstration data today comes from a handful of labs and a few large cities. East India adds people, workplaces, tools and objects that most models have never seen. Field operations here also cost less to run than in Delhi NCR or Bengaluru, so the same budget buys more hours of demonstrations, and no major data collector works in the region yet.
What to send us
- Your task list and how each episode should start and end.
- Your capture protocol, camera settings and sensor set.
- Any rules on operator variety, such as how many different people per task.
- Your metadata schema and delivery format, if you have one.
From demonstrations to policies
We focus on the part of the pipeline that is hardest to do from a lab: getting consistent, consented demonstrations out of real workplaces. Your team keeps control of annotation, model architecture and training. If you also need to test the resulting policy on hardware, we support on-site setup and operation for your robot evaluation runs.
Consent and compliance
Every participant gives written informed consent before recording, every site is covered by a written agreement, and faces can be blurred on request. Footage is used only for your model training and never resold. Read more about DPDP Act AI training data compliance.
Related data types
Imitation learning data overlaps closely with VLA training data, where demonstrations are paired with language, and humanoid robot data, where the focus is on manipulation.
Start with a pilot
Test our imitation learning data on your own tasks with the 50–100 hour pilot. You choose the task categories, send equipment and guidelines, and we deliver consented, quality-checked data within 2 weeks.