01 ↗
Is the source credible?
Can the supplier describe the recording's origin, collection method, date and context? Are usage and transfer rights documented?
DATA TRUST / PROVENANCE / VALIDATION
Know where your data comes from. Know what it contains.
Physical AI companies need more than hours of video. They need credible, traceable datasets, clear permitted uses and evidence that the recordings and annotations match the training requirements.
THE PROBLEM WE ADDRESS
Not every video is trustworthy training data.
Mixed suppliers, opaque collection methods, repeated clips, mislabelled actions and incomplete telemetry can reduce the value of physical-AI datasets. Our review approach separates four distinct questions.
02 ↗
Is the capture genuine?
Do available evidence, metadata and samples align with the claimed real-world capture, without unexplained editing or inconsistent streams?
03 ↗
Is the data technically usable?
Do files decode, timestamps align and important frames or sensor streams remain present, clear and nonduplicated?
04 ↗
Will it serve the task?
Do recorded actions, modalities, labels and outcomes match the customer's Physical AI learning objectives and agreed acceptance criteria?
THE VERIFICATION WORKFLOW
Evidence-backed checks from intake to delivery.
Verification plans can combine source documentation, automated checks and human review. The exact controls depend on the collection method, data type and available evidence.
01
Source provenance & rights
Review supplier identity, collection context, consent and permitted uses, capture descriptions, dataset ownership and evidence of the source chain.
02
Authenticity & anomaly screening
Review available metadata, continuity and consistency; flag unusual editing, repeated footage, suspicious patterns or mismatches between source claims and representative clips.
03
Technical & sensor integrity
Screen for corrupt files, black or static frames, duplicates, unexpected frame-rate variation, timestamp gaps, and missing or desynchronized camera/IMU/robot data.
04
Semantic & annotation QA
Check task relevance, action boundaries, object/interaction labels, metadata consistency and reviewer agreement against the agreed annotation specification.
05
Acceptance & exception review
Apply agreed thresholds and sample-based review to identify accepted, rejected, uncertain and rework-required records. Retain traceable review decisions.
06
Documented delivery
Where commissioned, prepare a dataset manifest with validated file inventory, modalities, checks performed, limitations, annotation version and outstanding exceptions.
PRACTICAL OUTPUTS
A clearer basis for accepting a dataset.
Instead of an unsupported claim that footage is “100% authentic,” agree in advance what evidence buyers need and what review records accompany delivery.
A ↗
Provenance record
Source and license documentation, capture description, custody notes and known evidence gaps.
B ↗
Quality-control report
Technical checks, exceptions, sampling methods, failure categories and remediation decisions.
C ↗
Acceptance-ready metadata
Dataset manifests, task labels, sensor details, review status and agreed output format.
Our approach: we assess and document confidence using available evidence. AI-generated/manipulated video detection and source authenticity cannot be guaranteed from video pixels alone. We do not represent these checks as independent certification or a guarantee of model performance.
MULTIMODAL ROBOTICS DATA
Built for EGO, UMI and teleoperation workflows.
Checks can be tailored to monocular and stereo RGB, synchronized camera views, IMU streams, hand/gripper states, timestamped robot actions and MCAP recordings where those data are captured and available.
Mono / stereo frame consistencyTimestamp synchronizationFrame integrityDuplicate detectionIMU / action stream checksAnnotation reviewSource documentationDataset manifests
Make data trust part of your requirements.
Send your dataset specification, sources and acceptance criteria. We can propose a scoped verification plan and reporting deliverables.