Guide
Personally Identifiable Information in Robotics Data
Short answer
Personally identifiable information (PII) in robotics data means anything in the footage that could identify a real person — mainly faces, licence plates, on-screen text, and spoken names or numbers picked up by egocentric and robot-mounted cameras. It is handled with automatic detection and irreversible blurring for un-consented bystanders, plus a documented consent chain for the actual participants, not by avoiding capture altogether.
Last reviewed:
What counts as PII in egocentric or robotics video?
A head- or wrist-mounted camera captures whatever is in front of it, which means PII shows up incidentally as often as it shows up on purpose. The categories that recur in egocentric and robot-learning footage are faces and heads (including bystanders, reflections, and partial or profile views), licence plates, on-screen or printed text such as phone screens, monitors, forms, badges and mail, and spoken identifiers — names, addresses, and numbers — in the audio track.
This guide is general information about how PII typically appears in robotics capture and how Firsthand handles it, not legal advice. What counts as personal data, and what obligations follow, depends on your jurisdiction and use case — confirm specifics with counsel.
How is each type of PII handled?
- Faces & heads
- Detected per frame, tracked across the clip, and irreversibly blurred into the pixels before delivery.
- Licence plates
- Detected the same way as faces and irreversibly blurred; in scope for outdoor and automotive captures.
- Screens & documents
- Legible screens, forms, badges and mail are detected and blurred so no account or identity data leaks.
- Spoken identifiers
- Names, addresses and numbers in the audio are scrubbed from the transcript and muted in the waveform.
- Consented participants
- The operator and anyone who signed a release stay in frame — the pass targets un-consented third parties, not the training signal.
- Sign-off
- A human reviewer confirms no un-consented person is identifiable before an episode clears, and the pass is recorded per episode in the batch data card.
Firsthand’s own de-identification pipeline. See /anonymization for the live before/after and the step-by-step pass.
Does blurring PII replace the need for consent?
No — they answer different questions and both are needed. De-identification (blurring faces, plates, screens, scrubbing speech) reduces the risk that an un-consented bystander is identifiable in the delivered footage. It does not substitute for a documented release from the actual participant performing the recorded activity, whose face and hands are the training signal and are deliberately left intact. A dataset needs both: bystander PII removed, and participant consent documented.
Why does the redaction need to be irreversible?
Reversible masking — a blur that could be switched off, or a separate un-redacted master kept alongside the delivered file �� leaves the underlying PII recoverable, which keeps the footage in scope as personal data rather than removing it from scope. Firsthand’s blur is baked into the delivered pixel data with no un-blurred master shipped and no key to reverse it, and the pass fails closed: if a reviewer is unsure a region is fully covered, the episode is held or discarded rather than downgraded and shipped.
How can a buyer verify PII was actually removed?
Each episode in the batch data card lists the de-identification pass, the detector version, and the reviewer who confirmed it, so the claim is checkable per file rather than taken on assurance. That record sits alongside the consent artefact IDs for the same episode, which is what makes the two separate checks — PII removed, participant consent documented — auditable together.
FAQ
Personally Identifiable Information in Robotics Data, answered.
01What counts as PII in robot training data?
Mainly faces and heads, licence plates, legible screens or documents, and spoken names, addresses or numbers in the audio track — anything in egocentric or robot-mounted footage that could identify a real person.
02Does removing PII hurt the training signal?
No, for manipulation and whole-body policies the signal is in the hands, objects, gaze and scene, not bystander faces. Only un-consented third parties are removed; the consented operator, their hands and their point of view stay untouched.
03Is blurred PII the same as removed PII?
Only if the blur is irreversible. A reversible mask or a separate un-redacted master left the identifying data recoverable. Firsthand’s blur is baked into the delivered pixels with no master and no key to reverse it.
04Does de-identification remove the need for participant consent?
No. De-identification targets un-consented bystanders in the frame. It does not replace a documented release from the actual participant performing the recorded activity — that is a separate, required check.
05How is PII handling verified on a delivered batch?
Each episode’s data card lists the de-identification pass, detector version, and the reviewer who confirmed it, so the claim can be checked per file rather than assumed.
Check it against the sample pack.
40 episodes across 4 environments, delivered in the exact schema these guides describe.