AuraOne / Human Data / Physical AI

Fill what your real-world Physical AI data is missing.

Generate verified coverage for the gaps that matter. Or turn existing real-world video into structured, reviewable data.

Tell us the gap, or let AuraOne help identify it. Synthetic Coverage and Machine Enrichment can be scoped independently or connected when the program needs both.

Scope a Physical AI program Compare the capabilities

Two independent capabilities

Use the capability the data problem requires.

Synthetic Coverage does not require Machine Enrichment. Machine Enrichment does not require Synthetic Coverage. Each capability must be contracted and production-qualified for its own program scope before use.

Synthetic Coverage

Generate what is missing.

Start from a buyer-defined or AuraOne-assisted gap. Generate, validate, review, and deliver an immutable coverage batch.

No enrichment prerequisite
Accepted Coverage Batch is the commercial delivery

Machine Enrichment

Understand what is already there.

Run multimodal inference over existing real episodes and route uncertainty through human review.

No synthetic generation prerequisite
Calibrated only for contracted output classes

Optional combined loop

Connect them when the evidence supports it.

Real video can be enriched, analyzed for a measurable gap, and then extended with Synthetic Coverage. The full loop is useful, but never a prerequisite for either capability.

Synthetic Coverage

Fill a known gap with verified examples.

Tell us what your dataset is missing, or ask AuraOne to help find an underrepresented condition. We generate the contracted coverage and prove what was delivered.

Program-scoped

Buyer-defined gap

Start with the failure you already know.

A coverage program can begin from a buyer brief. Automatic gap discovery is not required.

Low-light manipulation variants
Transparent or reflective containers
Recovery episodes after grasp failures
Target counts, formats, and acceptance thresholds

AuraOne-assisted gap

Use source evidence to refine the target.

When the program includes suitable source data and analysis, AuraOne can help identify measurable underrepresented conditions. The evidence and method stay attached to the gap.

Slice and condition coverage
Failure and recovery distribution
Scenario constraints and exclusions
Traceable source-to-gap rationale

Define the gapBuyer-defined or AuraOne-assisted
Specify the scenarioConditions, counts, formats, rights
Generate coverageApproved providers and recorded lineage
Validate and reviewQuality gates plus human decisions
Accept the batchImmutable dataset, manifest, and receipt

The Accepted Coverage Batch is the delivery.

AuraOne verifies the contracted coverage, quality, review, rights, provenance, and delivery evidence. The buyer then accepts the immutable batch. Customer model or robot evaluation can be attached later, but it does not block normal delivery or acceptance.

Machine Enrichment

Turn raw Physical AI video into reviewable structure.

Multimodal models can describe and segment existing episodes. Confidence thresholds route questionable outputs into an append-only human review record with preserved machine authorship.

Scoped independently

Semantic output

A machine-readable account of what happened.

Output classes are selected and calibrated for the contracted program; they are not implied for every dataset.

Episode and interval descriptions
Task phases and temporal event boundaries
Actions, objects, failures, and recoveries
Confidence, provider, model, and prompt lineage

Human control

Uncertain output becomes a review task.

Existing annotation, queue, persistence, and review machinery carries machine suggestions into an auditable human decision.

Program-specific confidence thresholds
Accept, correct, reject, or quarantine decisions
Append-only corrections with immutable machine authorship
Durable links back to the source episode

Ingest real episodesCustomer or Capture source data
Run multimodal inferenceDescriptions, phases, actions, objects
Route uncertaintyConfidence thresholds decide review
Human reviewAccept, correct, reject, or quarantine
Deliver enrichmentAnnotations with source and review lineage

Machine Enrichment stands on its own.

Synthetic Coverage is not required. Enrichment is included only when it is contracted and its providers, output classes, thresholds, and review routing are production-qualified for that program.

Evidence, not overclaiming

Four labels keep delivery separate from downstream impact.

The first two establish normal delivery. The second two are optional evidence that may arrive later. AuraOne does not guarantee that a coverage batch will improve a customer model or robot.

AURAONE_VERIFIED

AuraOne verifies the batch.

Coverage, quality, review, rights, provenance, format, and delivery evidence meet the contracted criteria.

CUSTOMER_ACCEPTED

The buyer accepts the delivery.

An authorized buyer accepts the immutable batch. This completes the normal commercial delivery path.

CUSTOMER_REPORTED

The customer reports downstream impact.

If the customer evaluates the batch in a model or robot, AuraOne can attach the reported result without presenting it as AuraOne-controlled proof.

AURAONE_CONTROLLED_EVALUATION

AuraOne runs a separately authorized evaluation.

Only a controlled, authorized environment can produce this evidence class. It is optional and outside normal batch completion.

What we capture

Teach machines how the world works.

Real demonstrations remain part of the Physical AI foundation. When a program needs new real-world evidence, trained operators can capture the motion, trajectory, force, and failure context — not just the picture.

Teleoperation

Trained operators drive your robot through the task. Every attempt is recorded.

Session, operator, and task record

Human demonstration

People perform the task themselves while cameras and wearables capture how.

Video, pose, and device metadata

Multimodal capture

Video, trajectory, and force streams recorded in sync.

Aligned streams with timestamps

Episode review

Reviewers watch every episode. Bad ones get reworked or thrown out.

Disposition, reason, and rework history

Technical validation

Dropped frames, clock drift, and sensor faults get caught before delivery.

Validation report per episode

Versioned delivery

Datasets ship with manifests, checksums, and a data card. Every version is reproducible.

Manifest, checksums, and data card

From one skill to a dataset you can train on.

Name the skill. We find the operators, run the capture, and review every episode before it reaches you.

Name the skill.

The task, the robot, and the environment
What counts as a successful attempt
How many episodes, and by when

Operators record. Reviewers check.

Trained operators run the task on approved capture profiles
Streams are recorded in sync and validated automatically
Reviewers accept, rework, or reject every episode

Receive reviewed episodes.

Video, trajectories, pose, and sensor streams
Review dispositions and failure labels
Manifests, checksums, and a data card

How a capture program runs

Four steps. Each one has an owner and a record you can inspect afterward.

01
Brief
Name the skill, the environment, and what counts as a pass.
Approved task brief
02
Consent
Rights are attached before the first frame. Operators see what they signed.
Consent receipt and license scope
03
Capture
Operators record on approved profiles. Every stream ties back to its task and device.
Session, device, and stream metadata
04
Review
Reviewers watch every episode. Failures get labeled, not deleted.
Disposition, reason, and failure labels
05
Deliver
Accepted episodes ship with manifests, checksums, and a data card.
Delivery manifest and acknowledgment

What you receive

What arrives in a delivery

Your contract sets the exact scope. Managed robotics programs collect and review human demonstrations, then coordinate a monitored handoff; queued transcode, copy, or webhook work is not represented as delivered.

Recordings: Video, depth, and audio streams recorded in syncAccepted capture sessions
Motion: Trajectories, joint states, and force readingsAligned sensor streams
Context: Robot, environment, and operator metadataSession record
Review: Accept and reject decisions, rework history, failure labelsEpisode review record
Package: Data card, delivery manifest, and per-file checksumsVersioned dataset release
Formats: Video as MP4 or MOV. Records as JSONL, CSV, Parquet or JSONCurrent supported delivery formats

What you receive

Three layers. One usable dataset.

Raw is never overwritten. Clean removes the dead time. Intelligence explains what happened.

Raw

The unchanged evidence, exactly as it was collected.

Original media and sensor artifacts
Immutable checksums and artifact identity
Device, camera, and capture metadata

Clean

The accepted task content, with the dead time removed.

Accepted task intervals and valid streams
Setup and interruptions excluded, each with a reason
Reproducible raw-to-clean mappings

Intelligence

The record that explains what happened, and whether it worked.

Task version, phase boundaries, and terminal state
One declared outcome per episode
Quality, alignment, and provenance evidence

Capture profiles

Two ways to record the work.

A phone scales to volume. Wearables catch the hands. Both carry their full device and kit lineage.

Smartphone Ego

First-person capture from an approved phone configuration. Scales to volume.

Approved phone configurations and mounting positions
Task briefs with initial and terminal state holds
Video, motion signals, and device metadata

Wearable Manipulation

Multi-view capture for tasks where hand visibility and timing matter.

Head view, one or two wrist views, optional chest view
Registered stream roles with full camera and kit lineage
Measured temporal alignment across every accepted view

Where we capture

Real rooms. Not a lab bench.

A kitchen at 7am does not look like a kitchen in a dataset. We record in the places your robot will actually work.

What the data lets you do

Each outcome depends on the scope your program agreed to.

What the data lets you do. Each outcome depends on the scope your program agreed to.
Outcome	Work	What you receive	Program fit
Train a manipulation policy	Collect successful demonstrations of one skill across varied objects and settings.	Episode count, task coverage, and accept rate.	Coverage applies to the environments and objects in the brief.
Build a failure set	Keep the attempts that went wrong and label how they went wrong.	Failure labels, source episodes, and reviewer rationale.	Reuse depends on your rights scope and retention policy.
Benchmark a robot release	Replay a fixed episode set against a new policy and compare outcomes.	Held-out episodes, scoring criteria, and per-episode results.	Results apply to the scoped episode set and criteria version.
Expand into a new environment	Run the same skill in homes, warehouses, or retail floors.	Environment coverage, location consent, and privacy zones.	Each site requires its own consent and privacy review.

Start with the gap

Bring the data problem, not a predetermined workflow.

Tell us what is missing, show us the real episodes that need structure, or scope both. We will define the applicable capability, evidence boundary, and acceptance criteria with you.