Egocentric
Head, chest, and smart-glass views for human manipulation workflows.
Physical AI DataOps
World Archive
We collect, enrich, and deliver real-world manipulation data for robotics labs, VLA models, and world-model teams.
200+
in-house operators
180s
review-ready clip units
Ego + Exo
synchronized capture
QA
human verified delivery
Anonymized proof
- Active conversations with frontier robotics and physical-AI teams.
- Bespoke capture programs for manufacturing, textile, retail, and assembly environments.
- No named-client claims, logos, or founder endorsements without explicit permission.
Capture system
Field data from the places robots actually fail.
World Archive builds task-specific capture programs across factories, workshops, textile floors, retail aisles, packing stations, and tool-use environments. The output is not a video reel; it is a traceable dataset package.
Exocentric
Fixed side, overhead, and wide scene cameras for world context.
Wrist Cameras
Close-range object contact, tool handling, and fine-grain assembly.
Depth
Sensor depth where available, with clearly marked monocular estimates.
IMU
Motion traces from hands, wrists, tools, or capture rigs.
Language
Task intent, spoken notes, sub-step captions, and failure annotations.
Enrichment
Every clip ships with machine-readable context.
Automated labeling gets the first pass. Human QA closes the loop. The deliverable is built for model teams that need synchronized media, annotations, provenance, and clip quality in one place.
Object Tracks
Boxes and persistent IDs for manipulated objects, tools, and fixtures.
Segmentation
Masks for hands, target objects, tools, and work surfaces.
21-Point Hands
Frame-level 2D hand landmarks, plus estimated 3D when capture permits.
Action Boundaries
Reach, grasp, insert, align, press, release, inspect, retry.
Depth And Pose
Camera pose, scene anchors, depth maps, and calibration metadata.
QA Reports
Blur, occlusion, PII, manipulation density, success, and rejection notes.
Sample dataset pack
A 180-second clip becomes a small robotics dataset.
Clients can review a compact clip, inspect overlays, and ingest structured labels. Full parent sessions stay available for licensing, but 180 seconds is the clean unit for QA, buyer review, and model-data sampling.
overlay preview00:01:42 / 00:03:00
hand: 0.94contactdepth: ok
manifest
raw
videos/shuttle_001_raw.mp4
180-second egocentric source clip
preview
videos/shuttle_001_overlay.mp4
tracks, masks, depth, and action boundaries
labels
annotations/episodes.jsonl
task, sub-step, success, failure, and timestamps
pose
annotations/hand_keypoints.jsonl
21-point hand landmarks with confidence scores
vision
annotations/objects_coco.json
boxes, masks, object IDs, and frame spans
Task domains
Built around manipulation, not generic footage.
We prioritize workflows with repeated contact, tool use, alignment, grasp correction, inspection, and failure recovery. These are the behaviors physical-AI systems need to observe before deployment.
◇Shuttlecock assembly
◇Trophy and prize assembly
◇Sewing machine workflows
◇Packaging and kitting
◇Retail shelf handling
◇Warehouse sorting
◇Light manufacturing
◇Workshop repair tasks
Pipeline
From field capture to model-ready delivery.
Our operating loop is intentionally narrow: collect the real task, package it cleanly, annotate the signals that matter, verify quality, and deliver in formats model teams can actually use.
01
Capture
Operators record real manipulation tasks with ego, exo, wrist, depth, or IMU rigs.
02
Clip
Long sessions are segmented into buyer-reviewable 180-second units with parent-session links.
03
Enrich
Models draft hand tracks, masks, object IDs, depth estimates, and temporal labels.
04
Verify
Human QA checks visibility, privacy, contact states, and task success before delivery.
05
Deliver
Teams receive MP4, overlay previews, JSONL, Parquet, masks, and QA scorecards.
QA scorecard
Trust is a data field, not a slide.
Every dataset package includes visibility, privacy, manipulation density, task success, and annotation QA so buyers can decide quickly.
Hand visibility
94%
Manipulation density
87%
PII cleared
Yes
Task success
Pass
Depth source
Sensor / estimated
Annotation QA
Human verified
Delivery formats
Raw signal and structured labels, side by side.
| Media | MP4 raw clips, synchronized overlay previews, parent-session references |
|---|---|
| Annotations | JSONL episodes, COCO-style objects, masks, hand landmarks |
| Metadata | Device, location class, operator, task, timestamps, consent, QA |
| Reports | Privacy review, rejection notes, manipulation density, checksums |
Founder-led scoping
Send us the robot task. We will scope the dataset around it.
Share the target behavior, environment, modality, annotation need, and delivery timeline. We will map the capture protocol and sample pack.
No unverified named-client claimsCommercial-rights awareSecure deliveryModel-ready artifacts