Technical integrity
Decode success, resolution, frame rate, exposure, blur, corruption, and sensor continuity.
VIDEO DATA CURATION / ROBOTICS
Know what is in your video data.
Docent turns raw robotics video into a measurable, queryable, model-ready dataset built to your quality specification.
ACCEPT71.4%
REVIEW08.7%
REJECT19.9%
ACTIVE RULETASK.COVERAGE ≥ TARGET
THE SPECIFICATION
DEFINE QUALITY BEFORE SELECTION
01
A clean dataset is not enough. It must be the right dataset.
Docent converts your intended training use into a testable quality contract. Every asset is evaluated against the same requirements, with traceable evidence for why it was kept, rejected, or reviewed.
Semantic fit
Tasks, actions, objects, environments, outcomes, and edge cases matched to your target domain.
Episode integrity
Complete temporal sequences with aligned observations, actions, state transitions, and outcomes.
Diversity + coverage
Measured representation across behaviors, scenes, embodiments, conditions, and failure modes.
Uniqueness
Exact, near, and semantic duplicates removed without erasing useful behavioral variation.
Annotation quality
Captions, labels, segments, and metadata checked for consistency, confidence, and leakage.
THE CURATION PIPELINE
MACHINE SCALE / HUMAN PRECISION
02
Inspect every asset. Spend human attention only where it changes the answer.
Ingest
Index media, metadata, sensors, provenance, and existing labels.
PROCESSUnderstand
Caption scenes, segment actions, track entities, and embed every episode.
PROCESSDiagnose
Detect defects, mismatches, duplicates, outliers, gaps, and label errors.
PROCESSCurate
Filter, rank, cluster, rebalance, and select against the quality contract.
PROCESSVerify
Escalate uncertain cases to targeted human review and acceptance testing.
PROCESSDeliver
Ship clean subsets, manifests, annotations, scores, and complete lineage.
OUTPUTTHE DELIVERY
DATA + EVIDENCE + LINEAGE
03
A dataset your researchers can trust and interrogate.
CURATED CORPUS
The selected videos and episodes, normalized to the requested schema.
DATA MANIFEST
Per-asset metadata, captions, embeddings, quality scores, and selection status.
QUALITY REPORT
Coverage, defect rates, duplicate reduction, review results, and acceptance evidence.
REPRODUCIBLE RULESET
The specification, thresholds, prompts, tests, and versioned curation decisions.
INPUT
UNSTRUCTURED VIDEO CORPUSCONTROL
CUSTOM QUALITY SPECIFICATIONOUTPUT
MODEL-READY DATA + ACCEPTANCE REPORTDOCENT / VIDEO DATA CURATION