Capture teams active in India
Request data →
§01   Hero

We send capture teams into working factories, kitchens, and warehouses and record what humans do — frame by frame, with synchronized RGB, depth, IMU, and force data.

For research teams training manipulation policies, VLA models, and long-horizon planners that need footage from real work, not staged demos.

Request data → See a sample clip ↓
§02
Sample / the video is the proof

Five environments, one operator's view.

RAMBHA · IN · CASHEW GRADING (RAW)
00:00  / 00:00
CLIP 01 / 05 · CASHEW GRADING
RGB1920×1080 · 31 fps
AUDIO48 kHz · ambient
Currently showingCashew grading (raw)
LocationRambha, India
ModalitiesRGB · Audio
RigHead-mounted, wide-angle
Clips5 of many — auto-rotating
§03
Method / five claims, five details

Good egocentric data isn't scraped. It's produced.

01

Operators are trained for 40+ hours before their first shift.

Not crowdsourced. Not Turkers. Named people on a payroll, recruited from inside the industries we capture — line cooks, warehouse pickers, machine operators. They know the work. They wear the rig through their normal shift.

Operator roster · excerptROS-v4.1
OP-142Cashew grading · 18mo312h
OP-088Line cook · 3yr244h
OP-211Warehouse pick · 7yr198h
OP-056Dairy milking · 11yr176h
OP-173Surgical scrub · 4yr142h
OP-219CNC setup · 2yr 94h
02

Every rig is calibrated on site, every morning.

Intrinsics, extrinsics, IMU bias — checked against a known target before the shift starts and logged with the capture. Every clip ships with its calibration file. If a lens drifts, we see it the same day, not three months into training.

Calibration · RIG-07 · 06:14 localCHK-00412
Reproj. error 0.21 pxPASS
03

We record full shifts, not staged tasks.

Six to ten hour continuous captures. The model sees the mistakes, the restarts, the boring parts, the cleanup. Staged demos give you a policy that works in a lab; shift captures give you one that survives a Tuesday.

Shift 04 · OP-142 · 07:00–15:0008h 00m
07:0009:0011:0013:0015:00
Productive capture Mistakes / restarts / cleanup
04

Temporal and spatial alignment ship with the data.

Hardware-synced at capture with a single clock source across cameras, IMU, and force sensors. Sub-millisecond timestamp tolerance, documented per clip. You don't align our streams — we do.

Sync trace · frame 0143218Δt @ t₀
RGB
+0.00 ms
DEPTH
+0.12 ms
IMU
+0.04 ms
FORCE
+0.18 ms
Guaranteed < 0.8 ms across stackWITHIN TOL
05

We work in environments most labs can't access.

Signed agreements with operating facilities. Insurance, on-site safety protocols, worker consent. A grad student can't walk into a surgical ward with a GoPro. We spend six months on paperwork so the data is both good and usable.

Site agreements · activeLEG-v3
Facility agreements signedPer site
Operator consent forms on fileAll captures
Data-use counsel on retainerIndia & US
On-site safety officersAll deployments
Liability coveragePer-site
§04
Environments / what's actually in the dataset

The honest list.

Environment Location Modalities Status
Cashew grading line Rambha, India RGB · Audio Available
FMCG facility Tangi, India RGB · Audio Available
Home — kitchen and household tasks Ludhiana, India RGB · Audio · Force · Tactile Available
Construction site Bhubaneswar, India RGB · Audio Capturing
Incense stick factory Ogalapada, India RGB · Audio Available
— and more on request We add new environments every quarter. Tell us what you need.
Modalities per site reflect current production capture Request a site via §08
§05
Spec / data you can request

Here's what ships. No forms first.

Modalities01
RGB1× or 2× head-mount
DepthStructured light / ToF
IMU6-axis, head + wrists
ForcePer-glove, 32 channels
AudioBinaural, optional
GazeOptional add-on
Formats02
VideoMP4/H.265, raw Bayer
DepthPLY, NPZ, 16-bit PNG
IMU/ForceParquet, HDF5
BundleROS bag (optional)
CalibrationOpenCV YAML per clip
MetadataJSON-LD per shift
Resolution & rate03
RGBup to 3840×2160 @ 60
Depth1280×720 @ 30
IMU1 kHz, 6-axis
Force500 Hz, 32 ch
Audio48 kHz, 24-bit
Sync Δt< 0.8 ms
Annotations04
Task seg.Hand-labeled, 2-wk t/a
Object bboxEvery N frames, 1-wk t/a
GraspStart/end, type, force
LanguageFirst-person, per step
QC pass2 reviewers per clip
FormatCOCO + custom JSON
Licensing05
ExclusivePer-environment, time-boxed
Non-excl.Available to all customers
Term12 / 24 / 36 mo
Model useTraining + deployment
RedistributionNot permitted
Operator consentBaked into terms
Delivery06
New env.6–10 wk incl. agreements
Additional hrs2–4 wk from an active site
Annotations+1–2 wk over capture
TransferS3 / GCS pull, signed URLs
Pilot20-hr sample, 2 wk
SLANamed point of contact
§06
Use cases / what labs do with this

Three kinds of teams buy from us.

CASE 01

Manipulation policy training

The force and tactile channels are the reason teams come to us. You can train a grasp policy from RGB alone; you can't train one that doesn't crush a ripe tomato. 32-channel per-glove force data, time-aligned to the frame.

CASE 02

VLA & foundation-model pretraining

First-person language instructions from the operator, step-segmented, paired with vision and proprioception. Usable as a pretraining corpus for vision-language-action models without a second labeling pass.

CASE 03

Teleop bootstrapping & planning

Long-horizon captures — full 8-hour shifts — let planners see the distribution of a real task, including failure modes. Useful for BC bootstrapping before switching to teleop or RL refinement.

§07
Earn with Northstar / operators

If you work in a kitchen, warehouse, or line — we pay you to wear a camera for your shift.

We're not looking for actors. We're looking for experienced workers who'll do their normal job wearing our rig. 40 hours of paid training up front. Pay by the captured hour, not the shift. You keep the right to opt out of specific clips for up to 30 days after capture.

We operate across India and are expanding. If your industry isn't on our site yet, tell us what you do — we add new environments every quarter.

Apply to operate →
§08
Contact / request data

Email us if you're serious. We'll send a pilot.

One person reads this inbox. We'll reply within two business days with either a pilot proposal or a clear no.

Read by a human. No newsletter.