Glossary
What Is a Production Data Pipeline for Robotics?
Last reviewed: September 2026
Short answer
A production data pipeline for robotics is a repeatable process that captures, synchronizes, labels, checks, anonymizes and packages training data against a written spec, so every batch is as trustworthy as the last. It contrasts with demo data, which is gathered once to show a capability and is rarely documented.
01
Why it matters
Model teams retrain continuously. If capture, sync or QA drifts between batches, a regression can come from the data rather than the model, and nobody can tell. A production pipeline fixes everything except what the team wants to vary, and keeps an unbroken record — consent, sync, rejects — for each episode from the rig to delivery.
02
How Firsthand handles it
Firsthand runs capture, calibration, PTP sync (1.12 ms median, 1.94 ms p99, 2.0 ms reject ceiling), labeling, QA, anonymization and packaging as one end-to-end process. Deliveries arrive in RLDS, LeRobot or WebDataset with a data card, reported in validated hours, under a perpetual, buyer-owned license.
03
Pipeline stages
| Stage | What is fixed |
|---|---|
| Capture | One rig: head-mounted global-shutter 4K, wrist cameras, stereo, depth, IMU |
| Sync | PTP with a 2.0 ms reject ceiling |
| Labeling | 3D hand pose and action segments |
| QA | A written accept/reject rule; validated hours only |
| Anonymization | Faces and plates blurred with human review |
| Packaging | RLDS, LeRobot or WebDataset plus a data card |
05
Frequently asked questions
- How is a production pipeline different from demo data?
- Demo data is gathered once; a production pipeline delivers repeatedly against a spec with measured sync, validated hours, consent records and a data card.
- What formats does it deliver?
- RLDS, LeRobot or WebDataset, matched to the training pipeline.
- Is a production pipeline slow?
- No. After a signed spec, first validated data arrives in a median of 48 hours where coverage is deep.