Robotforge.Start a pilot ↗
SERVICES / THE DATA LIFECYCLE

From physical work
to training-ready data.

A collection is only the beginning. Build a project around the stages your pipeline needs, with outputs and acceptance criteria agreed before work starts.

CAPTURE → REFINE → VALIDATE → DELIVER

Ten stages. One deliberate specification. Choose a focused service or scope a connected workflow.

01

Capture / Collection

Record the work, in context.

Collect through our own growing network of partner workplaces, with access arranged for each pilot. Plan and record raw demonstrations using fixed, worn or multi-view cameras. Depth, motion and tactile sensing are assessed against the task and capture hardware.

02

Cleaning / Anonymization

Reduce exposure of identifying information.

Scope face blurring, text or logo redaction, voice treatment and removal of location metadata while preserving the task signals the buyer needs.

03

Curation / Filtering

Separate usable data from raw footage.

Review recordings against agreed criteria for lighting, occlusion, completeness and file integrity. Exclude or flag clips that do not meet the collection specification.

04

Segmentation / Structuring

Turn continuous footage into task units.

Identify task boundaries and divide recordings into discrete episodes. Preserve context and align start and stop points with the buyer’s definition of a demonstration.

05

Annotation / Labeling

Describe what happened, and when.

Define labels for actions, objects, task phases and outcomes. Pose or keypoint annotation is separately scoped for suitable footage and specialist review.

06

Enrichment

Add useful derived layers.

Evaluate depth estimation, 3D pose, object masks and action-language pairs as optional derived layers. Suitability depends on source quality, tools and validation requirements.

07

Quality Assurance / Validation

Make quality review explicit.

Review annotation accuracy, dataset consistency and anonymization effectiveness against documented criteria. Privacy checks assess residual re-identification risk; they cannot guarantee its absence.

08

Licensing / Legal Packaging

Document the terms of use.

Scope participant consent records, provenance and commercial usage rights. Buyer-specific privacy and regulatory requirements, including GDPR or CCPA considerations where applicable, require qualified legal review.

09

Delivery / Format Conversion

Fit the dataset to your pipeline.

Agree codecs, naming, manifests and dataset schemas. ROS bags, RLDS or LeRobot-compatible packaging can be assessed where the source contains the required fields.

10

Storage / Hosting

Plan for datasets that keep evolving.

Scope ongoing hosting, dataset versions and access management for recurring collection or subscription delivery, where included in the project agreement.

Define the output.
Validate a sample. Then scale.

Services, sensors, derived layers and delivery formats are confirmed per project. Human video does not by itself contain robot actions, calibrated depth or ground-truth pose. Legal and privacy requirements need explicit review; no blanket compliance certification is implied.

Let’s define the right dataset.

Start with a focused brief, an agreed protocol and a sample you can evaluate.

Discuss a pilot ↗