What is manufacturing robotics data collection?
Manufacturing robotics data collection is the capture of demonstrations and sensor data that train robots to assemble, pick, tend machines, and inspect parts on real production lines. Unlike programming a robot point by point, it records how skilled work is actually done, including force, part variation, and rare defects, so learned policies and inspection models hold up at line speed.
Factories already run on robots. The IFR counted 542,000 industrial robots installed in 2024 and 4.66 million in operation worldwide. Most of them are programmed, not trained.
What's changing: the tasks left over, like cable routing, flexible parts, mixed-model kitting, and cosmetic inspection, are too variable to hand-program. They need learned models, and learned models need data recorded where the work happens.
Where manufacturing robots need better data
| Application | What the robot must learn | Data we collect | Services |
|---|---|---|---|
| Contact-rich assemblycable routing, connectors, snap fits | Alignment, insertion force, when a part is seated | Teleoperated demos with force-torque, tactile, and wrist cameras | Teleop, multimodal |
| Kitting and bin picking | Mixed parts, tangles, shiny and dark surfaces | 3D scans and RGB-D capture of real bins across part lots | Point cloud |
| Machine tending | Loading CNCs and presses, door and fixture handling | Teleoperated cycles on the real cell | Teleop |
| Visual inspection | Rare, subtle defects under changing light | Multi-angle capture with defect sourcing and staged variants | Multimodal, edge case |
| Humanoids on the line | Tote handling and assembly steps built for people | Real-robot teleop and egocentric capture of line workers | Teleop, multimodal |
The Cycle-Time Reality check: data recorded the way the line runs
Gamasome framework
Lab demonstrations run at the operator's pace. Lines run at takt time. The Cycle-Time Reality check makes sure a manufacturing dataset reflects the actual line, not a slowed-down version of it.
A policy trained at half speed is a policy for a line that doesn't exist.
Takt-speed demonstrations
Operators qualify at line pace before their episodes count, so the policy learns motions that fit the cycle.
Part and lot variation
Parts from different suppliers, lots, and tolerance bands are captured and tagged, because variation is where hand-programmed cells fail.
Rare defects, sourced on purpose
Defect examples are pulled from quality holds and scrap, and staged where safe, instead of waiting for them to appear.
Fixture and tooling drift
Fixture wear and tool changes are logged, so a policy's drop in success can be traced to the cell, not the model.
Capture without stopping the line
Sessions are scheduled around shifts and changeovers, with capture rigs that don't add downtime.
What a contact-rich assembly program looks like
Composite scenario: connector mating on a wire harness line (details generalized)
- Situation
- A team is training a robot to seat connectors on a wire harness, a task still done by hand.
- Problem
- The policy looked good in the lab but failed on the line: new harness lots were stiffer, and the robot couldn't tell a half-seated connector from a fully seated one.
- Solution
- Teleoperated demonstrations at line pace across three harness lots, with wrist force-torque and audio to capture the seating click, and tagged half-seated failures kept in the dataset.
- Outcome
- A dataset where 'seated' is measurable in force and sound, not just in images, plus a held-out harness lot to prove the policy generalizes.
For the plant's manufacturing engineer, the win is a cell that doesn't need re-teaching every time a supplier changes.
Common mistakes in manufacturing robot data
Recording in the lab, deploying on the line
Lighting, vibration, and pace differ. Capture on the real cell, or on a replica validated against it.
One supplier lot
Models overfit to one lot's stiffness, color, and finish. Tag and span lots from the start.
Waiting for defects to appear
Rare defects arrive too slowly to train on. Source them from quality holds and stage variants safely.
Vision only for insertion
Seating is often invisible. Add force and audio, as in our multimodal data collection.
Read our guide to Physical AI in manufacturing and smart factories, and see how we prove a cell before rollout in validation and testing.
Manufacturing robotics data FAQs
What is manufacturing robotics data collection?
It's the capture of demonstrations and sensor data that train robots to assemble, pick, tend machines, and inspect parts on real production lines, including force, part variation, and defect examples.
Why do factory robots need training data if they can be programmed?
Programming works for repeatable tasks. Tasks with high variation, such as flexible parts, cable routing, mixed kitting, and cosmetic inspection, are easier to learn from demonstrations and data than to hand-code.
Can you collect data without stopping our production line?
Yes. Sessions are scheduled around shifts and changeovers, and capture rigs are designed so they don't add downtime to the cell.
How do you get enough defect examples for inspection models?
We source defects from quality holds and scrap, stage safe variants, and capture them under the same lighting and angles as production, tagging each by defect type.
Do you record force data for assembly tasks?
Yes. Contact-rich demonstrations include wrist force-torque and, where useful, tactile and audio streams on a shared clock.
Can you support humanoids working on factory lines?
Yes. We run real-robot teleoperation and egocentric capture of line tasks for humanoid teams; see our humanoid industry page for details.