Robotics & manipulation
Demonstrations of dexterous, tool-heavy, two-handed work for imitation learning and policy pretraining.
Physical AI & robotics data
Egocentric construction video, labeled by the trades.
First-person footage of electricians, plumbers, framers, welders, and other skilled tradespeople doing real work on real job sites — with step timestamps, narration, and expert labels. Built for robotics, embodied AI, and world-model teams.
What it is
First-person video of complete, real tasks.
The long tail of physical manipulation — wiring a box, sweating a joint, hanging and finishing drywall, laying a weld bead — lives in the hands of skilled tradespeople, not on the open web. We capture it from their point of view.
- Egocentric (POV) video from head-mounted or chest cams, multi-cam on request
- 1080p at 30 fps or higher
- Whole tasks from setup to cleanup, not cherry-picked clips
- Spoken or written narration of intent while working
- Real job sites and real materials, filmed with site and owner consent
First-person trade capturePhysical AI
Who it’s for
Built for teams teaching machines to work in the physical world.
Embodied AI & world models
Long-horizon, real-environment video with action segmentation and intent narration.
Vision & perception
Tool, object, and hand-interaction labels on cluttered, real job-site scenes.
Trades covered
Coverage across the skilled trades.
How it works
Capture, consent, and labels done properly.
Capture
Vetted tradespeople record their own work first-person on real sites, to a capture spec agreed during scoping.
Consent & privacy
Contributors sign a rights agreement, site and owner consent is collected, and faces, bystanders, and PII are blurred before data moves downstream.
Expert labeling & QA
Practitioners from the trade annotate steps, actions, tools, objects, and technique notes, then multi-pass QA reviews every record.
What’s in a record
Every clip ships with the context a model needs.
- Step-level timestamps and action segmentation
- Narration of intent, transcribed
- Expert notes: technique, safety, common mistakes
- Tool, object, and hand-interaction labels
- Trade, task, and contributor experience metadata
- Privacy flags, consent records, and QA status
VideoMP4
LabelsJSON · JSONL · CSV · parquet · COCO
DeliveryS3, Hugging Face, or your bucket
DocsData card + schema
Every delivered recordYour schema
- Source
- Domain, task, modality, expert role & experience
- Content
- Media or text, with timestamps where relevant
- Labels
- Steps, actions, objects, intents, rubrics, scores
- Expert notes
- Rationale, technique, edge cases
- Privacy
- Face blur & PII redaction flags
- Consent
- Contributor agreement & site/owner consent
- QA
- Review passes and approval status
- Format
- JSON · JSONL · CSV · parquet · COCO
FAQ
Egocentric trades video: common questions
What is an egocentric construction video dataset?
First-person (POV) video recorded from a head- or chest-mounted camera worn by a working tradesperson, paired with step timestamps, narration, and expert labels describing what is being done and why.
Which trades do you cover?
Electrical, plumbing, framing, drywall, concrete, HVAC, roofing, welding, tiling, painting, carpentry, and more on request.
What are the capture specs?
Head-mounted or chest-mounted cameras, with multi-cam on request, recording at 1080p and 30 fps or higher, on real job sites with site and owner consent.
How is privacy handled?
Contributors sign a rights agreement, site and owner consent is collected where filming, and faces, bystanders, and PII are blurred or redacted before data moves downstream.
Who labels the footage?
Practitioners from the trade — not generic crowd workers — annotate steps, actions, tools, objects, and expert notes, followed by multi-pass QA.
How is the data delivered?
MP4 video with JSON, JSONL, CSV, parquet, or COCO labels, delivered to S3, a private Hugging Face repo, or your bucket, with a data card and schema docs.
Start with a sample
See the data before you commit.
Tell us what your model needs to learn. We’ll scope a pilot and send a sample set in your schema.