Real-world data for physical AI

Stereo vision data.
Stable, consistent calibration.

Synchronized left/right video captures real-world manipulation with stable, consistent stereo calibration, providing a consistent geometric reference for spatial perception and hand–object learning. Explore task-matched clips, annotations and QC records before scaling your training set.

Stereo vision data Stable, consistent calibration Hand & body pose Per-frame labels Segment annotations Privacy screened

Preview the data

130,000+ h

90,000+ h

25 / 137

Snapshot 14 Nov 2026, 08:25 UTC+8. Externally shown as ranges; exact counts and full sample clips are released under NDA.

Catalogue

Four data products, one capture pipeline

Every product below is cut from the same stereo session, so video, pose, labels and segments stay frame-aligned.

STEREO · 1920×1200

RAW RGB

Stereo RGB video

100% of accepted clips

Synchronised left/right eye pairs recorded head-mounted at 30 fps inside live workplaces. Raw, undistorted and time-aligned.

1920×120030 fpsStereo pairH.264 / H.265

Download sample

21 KEYPOINTS · HAND

POSE

Hand & body pose

Pose-labelled subset

21-keypoint hand pose and full-body skeletons tracked per frame, with visibility flags for occluded joints.

21 hand keypoints17 body joints30 fpsJSON / COCO-style

Download sample

digital scale

keyboard

PER-FRAME · 1 fps LABELS

VLM SCREENED

Per-frame semantics

137 scenes · 78 task types

Frame-level scene, task and object boxes, plus a VLM privacy screen for minors, bystander faces and on-screen personal data.

EN / 中文 labelsBounding + polygonPrivacy flags

Download sample

action segments · 5 of 12

SEGMENT LABELS

TIMELINE

Segment & action annotations

Clip-level timeline

Start/end timestamps for each action segment, with success, hesitation and rework states — ready for behaviour cloning and reward modelling.

Action segmentsSuccess / reworkTimeline JSON

Download sample

Composition of the sellable library

share of accepted clips

Stereo RGB videoevery clip

100%

Hand & body posepose-labelled

~62%

Per-frame semantics1 fps labels

~48%

Segment & action labelsclip timeline

~35%

Layers stack on the same clip, so coverage is nested rather than exclusive; per-layer inventories are released in the NDA data room.

Overlays shown above are annotation-visualisation previews rendered on real delivered clips. Sample files are 8–10 s excerpts; full-length clips ship with the NDA sample pack.

Scale & growth

Volume is compounding, week over week

Cumulative qualified hours from first capture on 6 Jul 2026, with the current run-rate extended to year-end.

Qualified hours, cumulative

actual → projected

Quality mix of qualified hours

Hard-difficulty hours carry the highest label value and the tightest QC bar.

Countries with delivered footage

measured production footprint

Only countries with accepted, QC-passed deliveries are marked. Country volumes are not shown and no regional comparison is published here.

Qualified hours by data layer

derived from measured layer coverage

Layer hours are the measured library total multiplied by the coverage of each annotation layer (100% / 62% / 48% / 35%); they are not separate captures.

Clip-level QC record

Each clip ships with its own inspection record

Scene and task tags, a billable-share figure, an issue scan down to the second, and a timeline that shows what was blocked, what was waived and what made the final cut. The sample below is the same record our QC team works in.

Clip QC record

DR-2026-1113-0428 Creative workshops · Craft workshop · Hard

QC passed · billable

Raw

29:14

Blocked interval

14 s

Evaluation evidence

QC + segments

Capture

Stereo · DR-H01S

L1 Clothing & Fashion L2 Handcraft workshop L3 Craft workshop Skill · Sewing & laundry No gloves Normal

VLM summary

Threads a needle with embroidery floss, sorts the coloured threads and hand-stitches the border of a shoe insole.

Issue scan

per second, over the full 29:14

Hands not visible0 s

Idle0 s

Staged action0 s

Phone use0 s

Motion blur0 s

Lens occlusion14 s · blocked

Passer-by facesflagged only

Frame download cutwaived

Timeline lanes

one rail per review pass

Filter pass

Filter reject

Task labels

Sub-task labels

Final cut

00:0009:4519:3029:14

11:20–11:34 lens occlusion — the only blocked stretch; the rest of the clip is clean

How a clip is judged

One rule set, applied second by second

Normal Blocking Waived (shown, not blocking) Informational only

Content and task

Hands not visible — both hands out of frame for more than 10 sblocking

Idle, staged action, phone useblocking

Task or scene does not match the declared L1 / L2 / L3blocking

Motion blur that hides the manipulationblocking

Human-to-human interactioninformational

Image and device

Exposure, light interference, flipped frame, corrupted frame, stutter or black screen, lens occlusionblocking

Frame download cutwaived

Privacy and compliance

Passer-by faces, minors, private data on screen, non-first-person view, crew face in shotinformational

Billing clip — clean footage, at least 30 s, never cut per task, no upper limit.

Final cut — the same clean footage, segmented on task or step boundaries, each piece 30 s to 10 min.

Grey means dropped. Privacy and human-interaction flags never block a clip; they are recorded and reported to the buyer.

Record structure follows the production QC schema; per-clip scores, verdicts and raw log files are released with the NDA sample pack.

Task cards

Every task ships as a card: clip, labels, QC record

Nine task families below cover most of the sellable stock. Each card is built from real delivered clips with pose overlay, per-frame labels and a QC score.

Task types and card volume

78 task types · ≈35,000 task cards

Counts are rounded to the nearest 100 cards.Other 48 task types ≈2,200 cards

Hours are rounded ranges of delivery-ready stock per task, not single-contract figures; exact inventory and clip IDs are released with the NDA sample pack.

Stereo vision data.
Stable, consistent calibration.