Robot foundation models are starving. We refine the fuel.

We record real work twice. Once as the human sees it: 360° egocentric takes, up to ten hours without a cut. Once as a robot would act it: UMI-style gripper demonstrations, 6-DoF pose and width on every frame. Consented. Verified. Ready to train.

Jump The refinery Specs Catalog FAQ

The constraint

The open web won’t teach robots to work.

No viewpoint, no labels

Internet video watches hands from the outside. No egocentric viewpoint, no hand pose, no action labels. More of it teaches a policy nothing new.

Free corpora are clips

The open datasets are staged minutes: a tabletop, one task, a cut. Nobody gives away whole consented shifts of real work. Nobody labels them for robots.

In-house is slow

A lab that builds its own facility waits one to two years for useful daily volume. The model is hungry now.

The scarce distribution is whole shifts of real work: packing lines, kitchens, laundries. Captured legally. Captured twice: as the human sees it, and as a robot would act it.

One take, five gates

Follow one take through the refinery.

  1. 01 / 05 · CAPTURE

    Captured, first-person.

    A collector wears an Insta360 X3 and works the shift: packing, cooking, washing, ironing, folding. The camera sees what the worker sees, up to ten hours without a cut. On the gripper line the same camera rides the Newlithic parallel gripper, UMI-style. The take carries its own action labels.

  2. 02 / 05 · CONSENT

    Consented, before the first frame.

    Every take carries its own consent artifact: DPDP-grade, revocable, one per take. Provenance is not paperwork added later. It is a field in the schema.

  3. 03 / 05 · QA GATE

    Verified, in seconds.

    Every take gets a verdict at the station: bank or retake. Four reason codes: sync, blur, occlusion, task validity. Nothing is filtered in silence. The rejects are logged.

  4. 04 / 05 · REFINEMENT

    Refined, to the action.

    Segments are seeded by the collector and refined by annotators. Egocentric takes gain per-frame MANO hand pose. Gripper takes carry 6-DoF TCP trajectories and continuous width, retargetable to any registered end-effector.

  5. 05 / 05 · EXPORT

    Delivered, model-ready.

    LeRobot v3, RLDS, or GR00T. The raw masters are kept forever. A future format is a re-export, not a re-collection.

The refinery, in numbers

79,000 Hours across real commercial task families: packaging, restaurant cooking, dishwashing, garment care
10 h Longest continuous take. Shift-scale, not clip-scale
100% Of takes carry a per-take consent artifact
2 Capture streams: 360° egocentric + UMI-style gripper
3 Export formats: LeRobot v3 / RLDS / GR00T

Proof · the data itself

Don’t take our word. Read a take.

take_0481 · tote_handoff INSTA360 X3 · 360° EGO SYNC 4.2 MS · OK MANO · 21 KPT / HAND
take_1127 · garment_fold INSTA360 X3 · 360° EGO SYNC 3.8 MS · OK MANO · 21 KPT / HAND
take_2093 · dish_rack_load UMI GRIPPER · X3 TCP 6-DOF · 30 HZ · OK APERTURE 38.4 MM
TAKE 01 / 03
Acceptance criteria, excerpt NL-SPEC-2.1
Camera
Insta360 X3, 360° egocentric
Take length
Up to 10 h continuous
Ego lane
MANO hand pose, per frame
Gripper lane
TCP 6-DoF + gripper width, per frame (UMI-style)
Segments
Subtask, collector-seeded, annotator-refined
QA
Reason-coded, rejects logged
Consent
Per-take artifact, DPDP Sec-6

The trial slice ships with the full data dictionary. Request it →

Buyer routing

Two streams, one refinery. Route yourself.

Pretraining · 360° egocentric

Whole shifts, as the human sees them.

A worn Insta360 X3, ten hours without a cut. Hands, contact, interruptions, recovery. MANO pose on every frame. The long horizon your world model has never seen. Catalog slices, non-exclusive.

Get the trial slice
Post-training · UMI-style gripper

Demonstrations a robot can act on.

The X3 rides the Newlithic parallel gripper. 6-DoF TCP pose and continuous width on every frame. Behavior-cloning grade. Retargetable to your end-effector. Commissioned on your ontology, judged by your gates.

Commission collection

Catalog

79,000 hours, browsable as inventory.

NL-PKG-0417 PackagingTote pack-out, handoff, palletize 11,400 360° EGOMANO Available
NL-PKG-0561 PackagingCarton erect, fill, seal 8,900 UMI GRIPPERTCP 6DOFAPERTURE Available
NL-KIT-0092 Restaurant kitchenPrep, cook, plate: full service window 9,600 360° EGOMANO Available
NL-DSH-0210 DishwashingRack load / unload, sort, shelve 6,100 UMI GRIPPERTCP 6DOFAPERTURE In refinement
NL-GRM-0138 Garment careIroning, folding, stacking 7,700 360° EGOMANO Available
NL-RCV-0033 RecoveryInterruptions, drops, error recovery 3,200 360° EGOMANO In refinement
Excerpt: 6 of 40+ catalog slices Browse the full catalog →

Commissioned collection

What a commissioned pilot looks like.

  1. Week 0

    Task ontology + acceptance gates co-signed.

    Your task families and your pass/fail criteria, fixed in writing before the first frame.

  2. Week 1

    Collection standard authored, stations calibrated.

    Mounts, sync checks, and station thresholds set against your gates.

  3. Weeks 2–4

    Weekly drops with QC manifests.

    Each drop ships with reason-coded accept and reject logs. You see the rejects, not only the survivors.

  4. Close

    Eval-sized slice, full provenance pack, raw preserved.

    Consent artifact IDs joined per episode. Camera-native masters kept for future re-export.

Coverage Task families your fleet never sees Whole-shift coverage of your target ontology, interruptions included.
Provenance Consent status unknown, unauditable Audit-ready per take. Artifact IDs your counsel can trace.
Formats Weeks of conversion scripts before training Native to your training stack: LeRobot v3, RLDS, or GR00T.

Governance

Consent provenance, auditable to the take.

Neutrality

No lab investors. Buyer vaults isolated. Codenames on every workforce screen. No one else’s pilot sees your ontology.

Consent

DPDP Sec-6 artifacts, one per take. Withdrawal propagates downstream and joins sellability at query time. A revoked take leaves the catalog.

Retention

Unredacted masters are default-deny, encrypted, deletion-attested. Every access is logged. Counsel may request the attestations.

For researchers and counsel

Questions, answered plainly.

What exactly is in a take?

Egocentric stream: 360° video from a worn Insta360 X3, up to 10 h continuous. Per-frame MANO hand pose. Subtask segments. Sync and QA metadata. The take’s consent artifact ID.

Gripper stream: the same camera on the Newlithic parallel gripper. Per-frame 6-DoF TCP pose and continuous width. Segments, QA metadata, consent artifact ID. Both ship with the full data dictionary.

How do you get robot-ready actions from handheld capture?

The X3 rides the gripper. TCP pose and width are recovered per frame from the take itself. No external rig. The demonstration and its labels come from one capture. That is what makes them behavior-cloning grade.

What license lanes exist?

Three lanes. An evaluation license: the trial slice. Non-exclusive commercial licenses on catalog slices. Commissioned collection, exclusive on your ontology. Raw footage is never resold.

Can our counsel audit the consent chain?

Yes. Every episode carries an artifact ID that resolves to a DPDP Sec-6 record. Signed releases sit in escrow, producible under audit terms. Withdrawal propagates to sellability at query time. What you license is what remains consented.

Which export formats do you support?

LeRobot v3, RLDS, and GR00T today. The raw camera-native masters are kept indefinitely. A future format is a re-export, not a re-collection.

How fast can a commissioned pilot start?

Week 0: the ontology and acceptance gates are co-signed. Week 1: first calibrated capture. Weeks 2–4: weekly drops with QC manifests. A scoping call fixes the calendar. Book one here.

No deck required

Start with the trial slice. Decide on the data, not the deck.

A free, license-clean slice of packaging and industrial egocentric takes. Full schema, QC manifest, and consent artifacts included.