datai.
Menu

Useful footage.
Clear expectations.

Human video is one part of a physical AI dataset. The right capture and annotation requirements depend on the laboratory’s training objective.

VIDEO + STRUCTURED ANALYSIS

Record the action.
Explain its context.

Submit original MP4 or WebM recordings with a task description, environment, viewpoint, declared frame rate and clear permission to capture and license the content.

NVIDIA Nemotron generates descriptions and timed actions. A media worker measures the original file; automatic gates keep incomplete, uncertain or flagged recordings out of purchasable hours. Analysis waits until the production service is connected.

The accepted package includes

  • Stable recording and contributor identifiers
  • Activity, task, objects, tools and viewpoint
  • Technical metadata and verified duration
  • Timed action annotations and outcomes
  • Contributor permission declarations and analysis provenance
  • File and upload-part checksum provenance
  • A versioned, non-exclusive AI license
Machine-generated, clearly identified. Every time segment is analyzed with sampled visual frames. Audio is not analyzed. Action boundaries are approximate, and a model cannot establish ownership or permissions from pixels. Download the model clip JSON Schema. This schema covers one model response, not the full recording package or order manifest.
Who adds which information?
StagePerson responsibleWhat they provide
Capture contextContributorOccupation and activities, task goal, environment, objects/tools, capture device and permissions.
Action segmentsNVIDIA Nemotron 3.5 Super VLStart/end timestamps, verb, object, goal, outcome and notes. For example: 0–12 seconds, “repot”, “plant”, success.
Automatic acceptanceMedia worker and versioned quality gatesSource hash, measured media, complete timeline coverage, valid labels, visual quality and model privacy flags. Uncertain results stay off sale; exceptional review remains available.
Advanced annotationsSpecialist annotators and tools, when commissionedBounding boxes, object tracks, segmentation masks, hand/object contacts, calibrated pose or depth. These are not included unless explicitly specified.
EXPORTS

Simple formats.
Traceable content.

Paid orders provide original video downloads, a JSONL manifest with nested action annotations, and a CSV index. A full-recording analysis JSON is included when the order snapshot has an active machine package. The order preserves the metadata and annotation snapshot used at purchase.

Manifest excerpt; times are milliseconds from the start of the recording.

{
  "schema_version": "datai.video.v1",
  "video_id": "…",
  "duration_ms": 12000,
  "annotations": [
    {
      "start_ms": 0,
      "end_ms": 12000,
      "verb": "repot",
      "object": "plant",
      "goal": "Repot the plant",
      "success": "uncertain",
      "note": ""
    }
  ]
}

Delivery formats and batch downloads

AN IMPORTANT DISTINCTION

Human actions
aren’t robot commands.

Human recordings do not contain measured robot states, actuator commands or calibrated robot trajectories. LeRobot and RLDS-style control datasets need additional real signals. We do not fabricate them or claim that a file conversion creates them.

For specialized schemas or annotation layers, agree on the requirement with datai before purchasing.

Discuss a data requirement

Permission is part of the recording.

Contributors must have permission to record and license the content, including the location, workplace, people and confidential information involved. Machine analysis does not establish these rights.

Read the marketplace terms and AI footage license and privacy notice before contributing or licensing footage.

Prepare a clear research brief.

Compare activities, capture conditions and annotation requirements before defining a collection.

Explore activity guidesRead the dataset buyer’s guideBrowse research resourcesDefine a dataset brief
Explore the marketplace