Glossary
Data Curation for Robot Pretraining
Last reviewed: October 2026
Short answer
Data curation for robot pretraining is the work of selecting, filtering, validating, balancing and documenting human and robot data before it enters a pretraining or mid-training mix. It decides which data is trusted and in what proportion, and it differs from capture, annotation, the mixture ratio and the validated hour, which is the unit that passes QA.
01
Why it matters
Volume alone does not make a pretraining set useful. Generalist’s GEN-0 post, reporting its own large-scale ablations, states that “data quality and diversity matters more than sheer volume, and that carefully constructed data mixtures can lead to different pretrained model characteristics.” EgoScale shows the other side: its egocentric estimates “are noisy due to unconstrained data collection,” yet “the scale and diversity of the data provide effective supervision.” What a buyer can control is what a dataset is checked against before it is trusted, and that is curation.
02
How Firsthand handles it
Firsthand delivers validated human egocentric capture data, collected fast, on demand and to a written spec. It does not curate datasets you already own, and this page does not describe a curation service. What it does is run QA gates at capture and before delivery — sync, calibration, coverage, labels, consent and anonymization — log failures in a reject log, bill per validated hour, and ship each batch with a data card in RLDS, LeRobot format or WebDataset under a buyer-owned commercial license. Those gates are checks a buyer’s own curation can lean on. Selecting, balancing and mixture design stay with the buyer.
03
What steps does curating robot training data involve?
Curation is a sequence of decisions about a candidate pool. Five steps recur: selecting what could enter, filtering out what fails a check, validating that what remains is trustworthy, balancing it across scenes, tasks, objects and people, and documenting the result so it can be audited.
The QA pages on this site map onto the filtering and validating steps. Cross-stream sync is held to a 2.0 ms ceiling and a breaching episode is rejected. Calibration must stay under 0.28 px RMS reprojection at an opening and a closing solve, with drift within tolerance. Coverage is checked against the spec, including failure cases. Frame-level labels are checked against calibration and sync, and consent and anonymization are verified independently of the technical signals.
Balancing follows the diversity axes in the pretraining quality checklist: scenes, tasks, objects and contributors, each checked separately from total hours. Documenting is the data card and the reject log.
04
How is data curation different from collection, labeling, data mixture and a validated hour?
Curation is the decision layer across its neighbouring terms: it takes captured and labeled data as input, produces the mixture ratio as an output, and uses validated hours as one measure of what passed.
05
What do GEN-0 and EgoScale state about data quality and filtering?
GEN-0’s post says its Table 1 compares models trained on 8 different pretraining datasets, collected with multiple data foundry partners, and that “having multiple data collection strategies at scale allows us to continually A/B test which data improves pretraining the most.” It does not publish filter rules or retention rates, and none are inferred here.
EgoScale says it applies “off-the-shelf SLAM and hand-pose estimation pipelines to recover camera motion and human hand trajectories.” Its Appendix C, titled “Details of In-the-wild Human Data Curation,” reports distributions across categories, environments, tasks and objects. For its smaller aligned mid-training set it states that human demonstrations use “the same camera configuration as the robot, with matched viewpoints and calibrated intrinsics.” These are each study’s own setups; this page states no result for Firsthand data.
06
What should buyers check when curating custom egocentric data?
Ask for the thresholds in the signed spec rather than a promise of quality: a sync ceiling, a calibration threshold, the required skill variations and failure-case coverage, and the consent and anonymization steps a batch must clear. Then ask for a reject log alongside the validated-hour total, and a data card recording jurisdictions of capture, the consent basis and the de-identification pass.
Check that the schema is consistent from batch to batch. Firsthand’s episode schema carries action_segments (start_ns, end_ns, verb, noun), contact_events, hand_joints and hand_visible, with all streams on one PTP clock domain and int64 nanosecond timestamps. Where contributor coverage is already deep, a signed spec can reach a validated first batch in a median of 48 hours, and a Spec Pilot is the normal starting point.
07
How does a validated hour relate to curation?
A validated hour is 60 minutes of footage that passed QA against the spec: sync within the ceiling, calibration solved and within drift tolerance, coverage cells filled and consent documented. It is the output of the filtering and validating steps, not the whole of curation. It says nothing about whether an hour is the right hour for your mixture, so two sets with the same validated hours can differ in diversity.
08
Curation steps, what each checks, and where this site states it
| Curation step | What it checks | Where the site states it |
|---|---|---|
| Select | Whether a candidate source fits the target skills, scenes and embodiment | Skill spec; the pretraining quality checklist, which phrases each check as a question for a public dataset or a provider |
| Filter: sync | Residual timing offset between streams against a ceiling | Cross-stream sync error entry: 2.0 ms ceiling, breaching episodes rejected |
| Filter: calibration | RMS reprojection at the opening and closing solve, plus drift | Rejection criteria guide: under 0.28 px |
| Validate: labels | hand_joints, hand_visible, contact_events and action_segments checked against calibration and sync | Rejection criteria guide |
| Validate: consent | A signed release per contributor; faces, license plates and screens blurred, with human review | Rejection criteria guide; consent chain entry |
| Balance | Diversity across scenes, tasks, objects and contributors, plus failure-case share | Pretraining data quality checklist |
| Document | Data card, reject log, and a consistent delivery format across batches | Data card entry; pretraining data quality checklist |
10
Data curation compared with neighbouring terms
| Term | What it is | How it differs from data curation |
|---|---|---|
| Data collection (capture) | Recording the data; Firsthand uses a head-mounted global-shutter camera plus two wrist cameras on one clock | Produces the candidates; curation decides which to trust |
| Data labeling | Adding annotations such as action segments and 3D hand pose | Labels are data; curation checks them against calibration and sync |
| Data mixture | The blend of sources and their proportions in a training run | An output of curation decisions; curation also covers what is excluded and how it is documented |
| Validated hour | 60 minutes of footage that passed QA against the spec | A unit that results from filtering and validating; it does not cover selecting or balancing |
11
Frequently asked questions
- What is data curation for robot pretraining?
- Selecting, filtering, validating, balancing and documenting human and robot data before it enters a pretraining or mid-training mix. It decides which data is trusted and in what proportion.
- Is data curation the same as data labeling?
- No. Labeling adds annotations such as action segments or hand pose. Curation decides which labeled data to trust, including whether labels hold up against calibration and sync.
- Is data curation the same as a data mixture?
- No. A data mixture is the blend of sources and their proportions in a training run. Curation is the set of decisions that feeds it, and it also covers what is excluded and how the result is documented.
- Does curation matter more than volume?
- GEN-0’s post states that, from its own large-scale ablations, data quality and diversity matters more than sheer volume. EgoScale reports that the scale and diversity of noisy, unconstrained data still provided effective supervision.
- What gets rejected before a custom egocentric dataset is delivered?
- Episodes that breach the sync ceiling, miss the calibration threshold, fall short of spec coverage, fail frame-level label checks, or lack documented consent and anonymization. They go to a reject log and are recaptured or dropped, not delivered unflagged.
- Does Firsthand curate my existing dataset?
- No. Firsthand delivers validated human egocentric capture data. It does not curate datasets you already own or sell a curation service.
- What should a custom spec say if I plan to curate the data myself?
- State the sync ceiling, the calibration threshold, the required skill variations and failure-case coverage, the consent and anonymization steps, and the delivery format. Ask for the reject log and data card with each batch.
Related terms