Supported outputs
Use the format your training stack already reads.
Exports are designed to preserve review decisions while staying portable. A team should be able to leave with the subset, the manifest, the data card, and the evidence needed to reproduce the decision.
LeRobot / OpenX
dataset folder
Reviewed subset references, generated data card, rights scope, accepted segments, and failure cluster manifest.
RLDS
trajectory shards
Trajectory-oriented export for TensorFlow/OpenX-style robotics pipelines.
HDF5
.h5
Large multi-modal bundle support for ALOHA, ACT, RoboMimic, and generic profiles.
Manifest only
.jsonl
A training-ready reference set without copying raw media through the app server.
MP4/MOV + data card
media bundle
Scoped video delivery with clip references, redaction events, consent receipts, license boundaries, and review state attached.
MCAP / ROS bag
telemetry bundle
Planned packaging for synchronized robotics telemetry, sensor streams, and time-series events.
Parquet / Zarr
analytics and arrays
Planned large-scale metadata, action labels, QA events, segment tables, and chunked arrays.
Large media
Exports stream; they do not buffer the dataset.
Large accepted datasets are packaged through asynchronous jobs. The exporter writes JSONL records and artifact manifests directly to object storage, attaches checksums and expiry windows, then hands the buyer a manifest or cloud delivery job. The app server should not read multi-GB artifacts back into memory to create a download.
Artifact manifest
checksum + expiry
Each artifact carries size, SHA-256, source segment lineage, consent/license references, storage class, and retention policy.
Cloud handoff
copy or signed link
Buyer delivery prefers object-store signed links or cloud-to-cloud copy with verification instead of app-server downloads.
Signed manifest
Every export explains itself.
The manifest names the reviewed subset, source dataset, review decisions, QA state, consent/license references, recall linkage, payload roles, checksums, and signature metadata.
{
"manifest_id": "mf_01J...",
"created_at": "2026-05-14T21:22:00Z",
"source_dataset": "sim_rollout_policy_b",
"consent_receipts": ["consent_01J..."],
"license_scopes": ["lic_01J..."],
"recall_linkage": "recall_01J...",
"bundle": {
"format": "rlds",
"sha256": "535553111de1...",
"episode_count": 40,
"size_bytes": 9842322224,
"expires_at": "2026-06-18T21:22:00Z"
},
"payloads": [
"robotics_reviewed_subset_manifest",
"robotics_episode_reference",
"robotics_failure_cluster",
"robotics_intervention_note",
"robotics_embodiment_card",
"robotics_consent_receipt",
"robotics_license_scope",
"robotics_data_card",
"robotics_sensor_qa_report"
],
"signature": "ED25519:VLPS..."
}Delivery manifest
Buyer delivery is object-level and recallable.
Delivery manifests list every object sent or linked to the buyer with destination, status, checksum, provider request id, and delivery time. Recall notifications are generated from that object list so an accepted clip can be traced from capture through export and delivery.
AuraOne Programs
The intake export is explicit.
The public app stages a local manifest before handoff. Actual upload requires user consent and configured intake credentials; the public source-linked repo does not hardcode production secrets or silently send private media anywhere.