AuraOne
Start a project

Export formats.

The shape your training stack already reads. Every export is checksummed. Every export ships with a signed manifest. The tuned weights stay with the customer.

Supported formats

HDF5

.h5

The default for large multi-modal bundles. Pose, trajectory, video frames, and metadata in one file.

Parquet

.parquet

Tabular slices. Metadata, scoring, and reviewer decisions. Ideal for filtering and analytics.

MCAP

.mcap

Time-ordered multi-modal logs. Preferred for replay and for teams already on ROS 2.

ROSbag

.bag

Legacy ROS 1 compatibility. Emitted on request.

Signed manifest

Every export ships with a signed JSON manifest listing the clips included, the review decisions that approved them, the safety rules version in effect, and the tuning-run identifier.

{
  "manifest_id": "mf_01J...",
  "tenant_id": "ten_01J...",
  "created_at": "2026-04-21T14:22:00Z",
  "bundle": {
    "format": "hdf5",
    "sha256": "9f1d...c3b2",
    "clip_count": 412
  },
  "safety_rules_version": "v4.2",
  "tuning_run_id": "run_01J...",
  "signature": "RS256 ..."
}

Weight retention

OpenVLA fine-tuning runs in the customer tenant. Tuned weights stay in the customer tenant. They're exported alongside the training record when the customer asks, and they aren't shared.

If AuraOne disappeared tomorrow, the customer still has the tuned checkpoint, the regression set, the decision history, the training signal, and the measurable curve. All five artifacts are portable.

See also