Custom action & semantic annotation
Define labels and delivery rules around your model. Choose an automated pipeline, a human labeling team or a hybrid workflow after a sample review.
03 / MAKE EVERY EPISODE COUNT
From raw recordings to a usable dataset.
Custom annotation through automated pipelines, human teams or hybrid review, with quality rules and traceable delivery set to your brief.
THE RIGHT STARTING POINT
Your tasks define the labels, not a fixed template. Agree the schema, ontology, review rules and output format, then select an automated pipeline, human labeling team or hybrid workflow against a representative pilot.
WHAT WE CAN SCOPE
Details that make the difference.
01 /
Ingest and validate
Inventory files, verify formats and identify corrupt frames, incomplete episodes, missing channels and timing issues.
02 /
Configure the labeling workflow
Specify which labels can be generated or checked automatically, which need human judgment and which require both. Test the proposed workflow on customer-approved samples.
03 /
Add semantic context
Describe objects, locations, goals and outcomes in clear language. Maintain a versioned ontology and representative examples.
04 /
Review and adjudicate
Calibrate reviewers against agreed samples. Track disagreements, corrections and exclusions with a customer-defined acceptance rubric.
05 /
Prepare the training schema
Map fields, units and coordinate conventions to LeRobot, RLDS or your own loader after a compatibility check.
06 /
Release with provenance
Deliver stable IDs, manifests, checksums and release notes. Agree access controls, redaction, retention and deletion requirements for the project.
DEFINE THE DELIVERABLEA clear handover.
A clear handover.
No missing context.
- Annotation guide and versioned ontology
- Pilot-validated automation, human staffing and review plan
- Input/output schema and transformation map
- Annotated episodes and reviewer issue log
- Batch quality and acceptance report
- Loader-validated sample in the selected format
- Manifest, checksums and known-limitations statement
FIND YOUR STARTING POINT
Full catalogExplore the offering.
Dataset processing & QA
Organize, validate and release datasets with reproducible manifests. Make exclusions, missing signals and transformations visible before training.
Training-format integration
Map episodes into your training schema, document coordinate conventions and test the loader before committing a full dataset conversion.
LET'S GET SPECIFICYour questions,
Your questions,
answered.
Can you customize annotation to our requirements?
Yes. Share your task definitions, ontology, target format and acceptance rules. We can scope an automated pipeline, a human labeling team or a hybrid process; a pilot determines the appropriate mix and review coverage.
Can you annotate our own recordings?
Yes, subject to a sample review, supported formats and confirmation of your right to share the recordings for processing.
Do you provide hand poses or 3D annotations?
These are additional scopes requiring appropriate sensor coverage and validation. We do not infer metric 3D quality from ordinary video alone.
Is a format conversion the same as retargeting?
No. Format conversion reorganizes available data. Retargeting to a different embodiment requires a separate mapping, feasibility and validation scope.
How is quality measured?
Choose the unit and thresholds in the pilot: episode validity, segment boundary tolerance, label agreement, missing data and review coverage. The delivery report uses those definitions.
YOUR NEXT MOVE
Let's build something