Products · Collection, Labeling, Evaluation

The delivery pipeline

Five standardized stages, each backed by a real system module — not a promise.

011–3 days

Scope confirmation

A guided requirement wizard captures data type, scale, labeling standard, budget and target delivery date. Upload your spec and sample data; we return a cycle and quote range based on historical projects.

02Per batch

Production

The task engine dispatches work by split rules. Collection links to site scheduling, equipment ledgers and field check-ins; labeling assigns work by skill tier, with AI pre-labeling to cut manual effort.

03Real time

Quality control

Automated QA plus human review: honeypot sampling, arbitration for disagreements, sampling rates adjusted by annotator credit. Failing batches go straight into rework.

04Sign-off

Acceptance & settlement

Sample-based acceptance, online sign-off or one-click rejection. Rejection reasons are categorized and re-dispatched automatically; settlement and revenue split follow acceptance.

05Ongoing

Multi-format delivery

Export in the agreed formats with a quality report, field definitions and a version number. Data lands in the asset repository with version, lineage and license management.

Line 1 · Model evaluation

Quality ratings and reward data for video generation, world models and recommendation systems.

Rating dimensions

Visual quality & physical consistency: sharpness, jitter, deformation, object penetration, motion continuity
Semantic & instruction alignment: prompt adherence, subject consistency, camera and composition
Temporal coherence: inter-frame stability, closed-loop actions, long-video logic
Content safety: policy-violating elements, sensitive content, minor protection risks
Subjective preference: multi-dimension scoring and pairwise comparison for reward models and DPO data

Capacity and quality assurance

Annotator admissionSkill tests + trial labeling

Agreement controlCross-labeling + arbitration

Honeypot sampling3%–15%, credit-based

Typical cycle10k items / 5 business days

SettlementT+3 after acceptance

Line 2 · Data collection

In-house, first-party collection across robotics, autonomous driving, live commerce and warehousing.

Robotics collection

Dual-arm, single-arm and mobile base embodiments; teleoperation, motion capture and multi-camera capture with joint states, force feedback and trajectory ground truth.

LeRobotHDF5MP4

Scene video collection

Live commerce, warehouse picking, retail and roadside scenes, with targeted long-tail capture and field crew scheduling.

MP4KITTICOCO

Multi-modal & point cloud

Synchronized multi-camera, LiDAR and microphone arrays, supporting 6DoF poses, 3D boxes and joint speech transcription.

PCDKITTIWAV

In-house capability	Description	System module
Project intake & work orders	Projectized management from requirement breakdown to collection plan	Collection ops
Site & field crew scheduling	Venue booking, shift planning, field check-in with location verification	Scheduling
Equipment asset ledger	Teleoperation rigs, mocap suits and multi-camera units with loan and repair records	Asset mgmt
On-site quality monitoring	Remote inspection and spot checks to correct issues before they become batch-wide rework	QA workbench
Automatic ingestion	Direct integration with the data asset repository, with automated de-identification and tagging	Asset repository
Teleoperation compatibility	Multi-model adaptation records so data transfers across embodiments	Compatibility layer

Line 3 · Data labeling

A multi-modal labeling and QA workbench that turns subjective standards into measurable output.

Multi-modal tooling

Image, video, point cloud, audio and text tools behind one interface, switching by task type with resumable uploads.

Custom templates & schemas

Define your own fields and rules or import an existing schema, with versioning and change-impact hints.

AI pre-labeling & pre-review

Model-assisted pre-labeling and pre-review cut manual effort; confidence scores feed sampling rates.

Human review workbench

Reviewer interface with batch operations, keyboard shortcuts and side-by-side diffing; feedback returns to the annotator in one click.

Multi-annotator arbitration

Disagreements are auto-adjudicated by rule first; samples above threshold enter a human arbitration queue with full audit trail.

Ground truth & honeypots

Known-answer items are injected randomly to measure accuracy, with sampling rates driven by project and annotator credit.

Automatic quality reports

First-pass rate, agreement statistics and error-type distribution generated per batch and shipped with the delivery.

Task-level pipeline tracking

Live boards showing how each batch is distributed across collection, cleaning, labeling, QA and delivery.

Delivery formats & compliance

Standard formats work out of the box; custom formats come with the conversion scripts.

Format	Use case	Modality	Standard delivery
LeRobot	Robot imitation learning / VLA training	Video + joint states	Yes
HDF5	Teleoperation and multi-modal time series	Video + force feedback	Yes
RLDS / TFDS	Large-scale policy training pipelines	Multi-modal	Yes
COCO / KITTI	Detection, segmentation, 3D boxes, trajectories	Image / point cloud	Yes
CSV / JSON / JSONL	Ratings, text and structured labels	Text / scores	Yes
PCD / PLY	Raw and labeled point clouds	Point cloud	Yes
Custom	Integration with in-house training frameworks	Any	On request

Compliance tags

Every dataset carries tags for personal information, de-identification status, cross-border eligibility and sensitive content — so you can judge usability before purchase.

De-identification & watermarking

Faces, plates and speech are automatically de-identified; delivered files carry traceable watermarks and full access logs.

Versioning & lineage

Unified metadata standards record data origin, capture devices, annotators and QA records, supporting rollback and incremental updates.