Standards

Mapping and conformance model for train-ready releases

We align collector inputs and buyer outputs through a versioned canonical schema, validation gates, and format adapters.

Mapping approach

Every release passes through one deterministic mapping path so teams can trace behavior across source, canonical, and buyer output layers.

Source formats

ROS bag2, MCAP, and annotation sources captured with provenance metadata.

Canonical core

Unified episode-centric representation with standardized field semantics.

Target adapters

Buyer-required outputs generated from one validated canonical graph.

FormatStageSupportNotes
ROS bag2IngestFullPreserve source timing + metadata
MCAPIngest/ExportFullPreferred multi-modal container
RLDS-styleCanonicalFullEpisode-first schema semantics
OpenLABELAnnotationAdapterOntology mapping with traceability
COCOExportAdapterBuyer-compatible export format

Integrity

Checksums, missing file checks, and source consistency.

Schema

Required field and type conformance to canonical definitions.

Temporal

Timestamp and synchronization threshold checks.

Completeness

Coverage checks across episodes and modalities.

Action validity

Action/state sequences validated against policy rules.

Privacy

Consent and policy class checks before release.

Versioning policy

Major releases capture breaking schema changes, minor releases add compatible fields, and patch releases document non-breaking corrections. Each release includes migration notes and linked conformance evidence.

Contact team