[ curation ]

Curation is where the data becomes usable

Raw egocentric, exocentric, and IMU streams arrive as separate, unsynchronized data sources. Egocentric video shows task intent and hand-object interaction but lacks spatial grounding and action labels. Exocentric views provide scene geometry and full-body motion but miss the operator's visual attention. IMU captures precise motion telemetry but is meaningless without visual context. Curation synchronizes all three streams to a common timebase, cross-validates what's actually happening across modalities, recovers action labels by fusing pose estimates from video with IMU-derived motion, and produces integrated training data for embodied models — VLA and WAMs — that learn how perception and motion connect in the real world.

The pipeline

01Collect
demonstrations captured across real environments in several formats
02Clean
drop corrupt, redundant, and low-information segments
03Synchronize
align every sensor stream to a common clock
04Segment
cut continuous capture into discrete task episodes
05Annotate
actions, objects, language, and outcomes
06Quality control
automated checks plus expert human review
07Package
versioned, calibrated, training-ready datasets

1
Collect
demonstrations captured across real environments in several formats
2
Clean
drop corrupt, redundant, and low-information segments
3
Synchronize
align every sensor stream to a common clock
4
Segment
cut continuous capture into discrete task episodes
5
Annotate
actions, objects, language, and outcomes
6
Quality control
automated checks plus expert human review
7
Package
versioned, calibrated, training-ready datasets

Keeping datasets healthy

Failures are kept, not filtered

A training model cannot rely solely on clean episodes, as mistakes in the real world are common and robots must be capable of recovering from failure. Drops, slips, spills, and subsequent corrective actions are intentionally retained and labeled in datasets as these failure scenarios provide important examples for a policy to learn robust error recovery.

Versioned and reproducible

Every dataset version is tracked and archived to ensure experiments remain reproducible. If a dataset update causes performance regressions, teams can revert to a previously validated snapshot.

Human-in-the-loop review

Automated checks catch corruption and drift. Expert reviewers catch the edge cases, the ambiguous interactions, and the safety-critical labels that automation still gets wrong.