Collection & quality

Clear requirements. Controlled collection. Verifiable quality.

We turn a robotics data brief into an agreed protocol, a representative pilot and traceable production batches. Every release is reviewed against written technical, task, metadata and sourcing requirements.

Discuss a pilot View data samples

Project workflow

From brief to accepted delivery.

A compact five-stage process keeps collection, QA and customer review aligned without making the project unnecessarily complex.

01 / SCOPE

Define the target behavior

Confirm tasks, embodiment, camera views, sensors, environments, labels, volume and intended model use.

02 / PROTOCOL

Lock collection rules

Document equipment, capture settings, instructions, file naming, synchronization and pass/fail rules.

03 / PILOT

Validate a small batch

Test representative tasks and edge cases, review ingestion compatibility and refine the protocol.

04 / PRODUCTION

Collect in controlled batches

Track collectors, sites, equipment and protocol versions while production expands.

05 / RELEASE

Review and deliver

Accepted media, metadata, QC results and supporting documents are reconciled before release.

Quality checks

Four areas reviewed on every project.

The exact thresholds are agreed in the statement of work. Checks are adapted to the data type rather than applied as a generic checklist.

Capture integrity

TECHNICAL

Confirm that required streams are usable and aligned with the agreed capture specification.

Required viewpoints and sensors present

Resolution, frame rate and file format checks

Timestamp, synchronization and dropped-stream checks

Task validity

BEHAVIOR

Verify that each episode captures the intended task, outcome and relevant interaction states.

Correct start state, sequence and completion state

Natural object handling and valid environment setup

Failure, intervention and recovery marked where required

Metadata consistency

STRUCTURE

Ensure media, episode records, labels and review status connect through stable identifiers.

Consistent episode and file IDs

Required task, object and event fields complete

QC status and rejection reason codes recorded

Rights and source records

PROVENANCE

Keep project-level evidence for the collection source, participant permission and permitted use.

Contributor consent or authorization records

Collection-location permission where needed

Supplier, subcontractor and licensing details documented

Responsible sourcing

Verification matched to the project risk.

Not every collection requires the same identity checks. We agree the necessary controls before onboarding collectors, suppliers or locations, while avoiding unnecessary personal data in the training-data package.

US-first, global-ready: projects can add customer-specific privacy, security and supplier requirements for the United States, Europe/EEA or India. Final obligations depend on the collection location, data content, intended use and contract.

Company and supplier KYBConfirm the contracting entity, delivery partners, relevant business details and approved subcontractors.

Contributor verificationIdentity, age, eligibility and region checks can be applied where required by the project.

Consent and noticeExplain the collection purpose, intended use and applicable participation terms before data is submitted.

Site authorizationConfirm that recording is permitted in homes, workplaces or commercial environments used for collection.

Access and minimizationRestrict sensitive onboarding records and keep them separate from general model-training files.

Change controlDocument material changes to suppliers, equipment, protocol, geography or delivery schema.

Delivery evidence

What accompanies an accepted batch.

The documentation package is scaled to the project. Restricted identity records are not mixed into the normal training-data delivery.

SOW

Requirements and quality planTasks, modalities, environments, thresholds, review method and change process.

MAN

Episode manifestStable IDs, stream relationships, timestamps, metadata fields and schema version.

QCR

Batch QC reportSubmitted, accepted, rejected and reworked quantities with issue categories.

SRC

Source and rights summaryCollection source, contributor and site controls, intended use and license basis.

CFG

Capture configurationEquipment, camera placement, settings, synchronization method and protocol version.

Start with a small batch and agree quality before scale.

Request a pilot View samples