Semantic annotation
Describe scenes, tasks, actions, objects, states, attributes, and interactions.
Annotation & Quality Assurance
Annotation designed around how embodied models learn
We start with the training objective, then turn model requirements into annotation standards that are executable, reviewable, and scalable.
01
Expert annotation capabilities
Task decomposition
Break long sequences into tasks, subtasks, actions, and atomic actions.
Object bounding boxes
Annotate manipulated objects, targets, tools, containers, and key context.
Hand keypoints
Annotate hand structure, pose, grasp state, and hand–object interaction.
02
Data quality inspection
- Missing, corrupted, duplicated, and anomalous raw data
- Label conflicts, omissions, false positives, and inconsistency
- Action boundaries, temporal alignment, and task logic
- Batch evaluation, issue return, correction, and re-verification
03
VLM + human expertise
VLMs generate candidate semantics, task phases, and risk signals. Experts confirm, correct, resolve boundary cases, and own final quality.
04
Closed-loop quality
- 01Specification
- 02Pilot
- 03Production
- 04Intelligent pre-check
- 05Human review
- 06Correction
- 07Batch acceptance
FAQ
Frequently asked questions
Can KeenTruth follow an existing specification?
Yes. We validate completeness and sample coverage, then calibrate execution through a pilot.
What if we do not have a complete specification?
We can help define label schemas and boundary rules from model goals, sample data, and acceptance criteria.
Do VLMs fully replace human annotation?
No. VLMs improve pre-annotation and risk triage; experts remain responsible for complex judgment and final quality.
Can you inspect another vendor’s annotations?
Yes. We can independently inspect third-party output and provide issue lists and quality statistics.
Contact
Move raw embodied data into training faster.
Tell us about your data, annotation requirements, and quality expectations.