ROBOT TRAINING DATA

Factories pay us to collect robot training data

First-person industrial manipulation, captured on our own hardware inside working factories — not staged labs, not recruited participants.

EGOCENTRIC RGB · WRIST IMU · HAND POSE · CYCLE SEGMENTATION
oven assembly
panel screw-driving · egocentric
01
position the power driver over the panel
0:00.00–0:03.00 · position_driver
02
drive screws into the side panel
0:03.00–0:06.00 · drive_screws
03
lift and reposition the driver
0:06.00–0:08.00 · reposition
04
reach down and align the next panel section
0:08.00–0:11.00 · align_panel
05
drive screws into the overhead panel
0:11.00–0:14.00 · drive_screws
06
continue driving screws along the panel
0:14.00–0:18.00 · drive_screws
07
reposition and align the next fastener
0:18.00–0:20.00 · reposition
08
drive screws into the panel edge
0:20.00–0:23.00 · drive_screws
09
move across to the black bracket
0:23.00–0:26.00 · move_to_bracket
10
drive screws into the black bracket
0:26.00–0:28.60 · drive_screws
1,500 h/week capacity30 facilitiesFirst contract — Encord (YC W21)
THE GAP

Robot foundation models are not compute-constrained. They're data-constrained.

Factories are constrained by their manual labor's efficiency.

HOW IT WORKS

Everyone else pays to create manipulation data. We get paid on both ends of the same footage.

FACTORYworking production line
NUMENORhardware · pipeline · rights
AI LABrobot foundation model
FACTORY EFFICIENCY PLATFORM
numenor / overview
Numenor operations platform — overall efficiency dashboard
WHAT THE FACTORY PAYS FOR — EFFICIENCY, LOSSES, RECOVERABLE HOURS← OPEN →
ROBOTICS DATA PLATFORM
numenor / dataset / appliance-assembly
Numenor data explorer — annotated episode with hand pose and action segmentation
WHAT THE AI LAB BUYS — ANNOTATED EPISODES, READY TO LOAD← OPEN →
$ — money in, from both sides⇢ — what we ship outCAPTURE PAID FOR BEFORE THE DATA IS SOLD
WHAT YOU GET

Capture spec

Every delivered hour ships to this exact specification — the sensors it was recorded with, the labels on every frame, the format it arrives in, and the rights that come with it.

DeviceNumenor NMR-01, in-house wearable capture unit
ModalityEgocentric RGB, head-mounted
Resolution1920x1080 @ 30 fps
Wrist telemetry6-axis IMU, 200 Hz, hardware-synced
Hand pose21-keypoint, per hand, per frame
SegmentationCycle boundaries, action units
LabelsVerb-object ontology
Delivery formatRLDS / LeRobot-compatible episodes
Throughput1,500 hours/week sustained
LicensingPerpetual, non-exclusive; exclusive available

Sample manifest

What buyers inspect before purchase. Load the sample, align frames to actions, find episode boundaries, verify rights — without asking us a single question.

numenor / dataset / sample_manifest

episode_id · task_instruction · facility_id · operator_session_id

observation_streams · timestamps · cycle_boundaries · success_label

consent_id · license · provenance

Download the datasheet (PDF) →
THE DEVICE

We build our own capture units.

Off-the-shelf action cameras weren't built for an eight-hour shift on a production floor. Ours are: a 1 kg wearable unit, 8 hours of runtime — a full shift — hardware timestamp sync between video and IMU, and a mount that survives a factory and doesn't slow the operator down.

That means we control unit cost, sensor configuration, and what a new capture site costs to open. A competitor sourcing consumer hardware controls none of it.

EGO RGB · 30 FPSIMU 200 Hz · HW-SYNC1 KG UNIT8 H RUNTIME · FULL SHIFTFACTORY-PROOF MOUNT
NMR-01 head-mounted capture rigHEAD RIG · EGO RGB + IMU
NMR-01 stereo capture capSTEREO CAP · DUAL RGB
NMR-01 head-strap capture unitHEAD STRAP · EGO RGB + IMU
NMR-01 depth capture capDEPTH CAP · RGB-D
NMR-01 wrist capture unitWRIST UNIT · RGB + IMU
NMR-01 wearable compute packCOMPUTE PACK · FULL-SHIFT BATTERY
NMR-01 CAPTURE FAMILY — BUILT IN-HOUSE
//PIPELINE — RAW FOOTAGE TO DELIVERED EPISODES, ALL OF IT OURS
01CAPTURE02HW-SYNC03CYCLE SEGMENTATION04HAND POSE05ONTOLOGY LABELS06RLDS EPISODES
COVERAGE

Where the data comes from

Live in Uzbekistan today. Each new region opens with the same playbook and the same devices, run by our own team.

6,000
hours of capture per month, live today
//PROCESSES
Six process families in capture: assembly, metalworking, sewing, filling, manual installation, warehouse picking.
UzbekistanLIVE

30 facilities · 6,000 hours/month. Assembly, metalworking, sewing, filling and packaging, manual installation, warehouse picking.

Central Asia2026

Opens on the same playbook, with the same NMR-01 fleet — run by our own team, not partners.

MENA2027

Aerospace MRO, shipbuilding, offshore — the rare processes that price highest.

— Rare processes price higher. An hour of aircraft engine MRO is not an hour of packaging, and the labs know it.

RIGHTS & PROVENANCE

Consent artifacts ship with the data. They are what makes the footage trainable, commercially.

//01ConsentEvery operator signs an individual, revocable release.
//02RightsCapture rights assigned to Numenor in writing by each facility. We license; they don’t.
//03AnonymizationFaces and identifying marks removed on request.
//04IPProcess-sensitive segments excluded per facility NDA.
//05ProvenancePer-clip chain of custody: facility, date, consent ID.
//06JurisdictionUzbekistan and UAE only. No sanctioned jurisdictions.
WHY THIS IS HARD TO COPY

Three moats, in order of strength.

//01
Not staged capture.

Most egocentric data on the market is recruited participants performing tasks for a camera, in homes and mock warehouses. Ours is people doing their actual job on a line that has to hit output today. Different distribution, real failure modes.

//02
The whole stack is ours.

We build the capture hardware, we run the pipeline from raw footage to delivered episodes, and we hold the factory relationships. Nobody in this business owns all three, which is why their cost per hour is set by vendors and ours isn’t.

//03
The factory pays.

Our competitors’ unit economics start negative. Ours start positive, so we can scale capture ahead of data revenue.

STAGED CAPTURE
NUMENOR CAPTURE
Who is recorded
Recruited participants performing for a camera
Operators doing their actual job
Where
Homes and mock warehouses
A line that has to hit output today
Task selection
Chosen by whoever pays for the rig
The factory’s actual process
Failure modes
Staged and clean
Real, in distribution
Rights
No consent chain
Consent artifacts ship with the data
TEAM

Three founders. Hardware, software, and the factory floor.

Islam Akramov
Islam Akramov
Co-founder, CEO

Factory access, data licensing, and the deal that pays for capture.

LINKEDIN ↗
Nusratullo Akhunov
Nusratullo Akhunov
Co-founder, Software

Built the entire pipeline: ingest, sync, hand pose, cycle segmentation, annotation, delivery.

LINKEDIN ↗
Rikhsitilla Rakhimjonov
Rikhsitilla Rakhimjonov
Co-founder, Hardware

Built the NMR-01 capture unit from scratch.

LINKEDIN ↗
FAQ

Questions buyers ask.

Anything else?
WE REPLY WITHIN 24 HOURS
Write to data@numenor.world →

Default licensing is perpetual and non-exclusive. Exclusive licensing is available per process or per facility — priced accordingly.

RLDS / LeRobot-compatible episodes: egocentric RGB at 1920x1080 @ 30 fps, 6-axis wrist IMU at 200 Hz hardware-synced, 21-keypoint hand pose per frame, cycle boundaries, and verb-object action labels — plus a manifest with consent_id, license, and provenance for every clip.

Yes — that is what booking capacity means. We sustain 1,500 hours/week across six process families today; rare processes price higher.

Every operator signs an individual, revocable release; facilities assign capture rights to Numenor in writing. Faces are removed on request, process-sensitive segments are excluded per facility NDA, and every clip carries its consent ID.

Load the sample: align frames to actions, find episode boundaries, and verify the rights chain — without asking us a single question.

Because the factory already paid for the capture. Teleoperation rigs run $30–100 per hour before margin; our capture is funded by the process-analysis contract.

ROBOT TRAINING DATA

The long tail of manipulation data is sitting on factory floors. We're already recording it.

data@numenor.world