Process

How we collect data

From brief to versioned delivery.

Workflow

01
Discovery & dataset specification
We document embodiment, sensors, task definitions, success/failure criteria, target episode counts, environments, exclusivity, NDA, and delivery format before capture begins.
02
Hardware setup & calibration
Standardized kits for egocentric and multi-camera work: RGB and optional depth, IMU when required, calibration routines, and sync verification. Controller or robot telemetry can be included when available.
03
Capture protocols
Versioned task scripts. Operators follow defined coverage (viewpoint, object sets, lighting bands). Deviations are logged so researchers understand distribution shift.
04
Annotation standards
Action segmentation, object/scene labels, optional language instructions, safety/quality flags. Annotator QA rejection rates are tracked. Taxonomy is versioned with the release.
05
Validation methods
Automated checks for blur, occlusion, missing frames, and sensor desync; human review sampling; schema conformance against the agreed manifest; per-episode quality scores shipped with the release.
06
Packaging & delivery
Immutable versioned bundles, provenance notes, QA summary, and loader documentation. Formats follow the contract: LeRobot-oriented packages, RLDS, HDF5, or MP4 with JSON sidecars.

Supported formats

Packaged for your training stack

Prefer LeRobot loaders? See LeRobot data collection.

Security & consent