Commissioned first-person manipulation data, captured to your spec by a verified contributor network: VLA-annotated, hand-pose enriched, PII-scrubbed, provenance-hashed, and delivered format-native. Loader-validated on every export.
Pilot bounties start at $300: you write the spec and pay only for clips that pass your acceptance criteria.
Web video is exhausted. Sim-to-real has a ceiling. In-house collection produces one kitchen, one lighting condition, one pair of hands. We operate the layer in between: genuine homes, permissioned worksites, and verified expertise.
Kitchens, laundry, cleaning, organization, tool use: high-volume egocentric footage across dozens of unique environments per order, segmented and language-labeled.
HVAC, electrical, plumbing, machining, commercial kitchens: licensed professionals on permissioned worksites, with domain-accurate action vocabulary no crowd platform can produce. The footage nobody else can source.
Collected-after-date footage held out of every training delivery, enforced in an append-only registry rather than promised in a contract. Contamination-free by construction, provable clip by clip.
Send us your existing recordings. We return temporal action segments, language instructions, success flags, hand-pose tracks, and PII scrubbing, all in your training format and with a validation report.
Raw footage is not training data. Seven automated stages stand between a contributor's phone and your training loop, and the last one proves the dataset actually trains.
Contributors record against your spec with framing guidance and per-task checklists. Technical properties measured on upload: resolution, frame rate, stability, exposure.
Vision-model scoring against your acceptance criteria (task completion, hand visibility, framing, spec conformance), with confidence bands and human review on the margin.
Every frame face-scanned; detections blurred with temporal smoothing. Face-clear clips ship untouched with their scan manifest. Originals segregated, never modified.
Temporal action segmentation, per-segment language instructions, object references, success flags: generated by a multi-stage vision pipeline and gated by dual-model consensus, with disagreements flagged, never silently shipped.
21 keypoints per hand, per frame, with per-keypoint confidence and explicit occlusion handling: gloves and tool-occluded moments flagged, never hallucinated. Commercially-licensed extractor, versioned and regenerable.
Every accepted clip hashed with its consent record at acceptance. Contributor payouts carry the clip hash in the on-chain memo. The rights chain is independently auditable, clip by clip.
LeRobot v3.0 (v2.1 on request), RLDS, HDF5/robomimic, delivered as episodes, not folders. Full data card: label provenance breakdown, pose coverage, consent references, model manifests.
Before you ever see it, every delivery loads in the current LeRobot release and trains a reference policy on a held-out split. You receive the loss curve. If it doesn't train, it doesn't ship.
No minimums, no lock-in, no enterprise sales cycle. Every rung pays only for footage that passes your acceptance criteria.
A sample of the full stack: annotated, pose-enriched, hashed, format-native, with the smoke-test report attached. Load it in a training run this week and judge the pipeline on your own metrics.
Request the slice →Send your taxonomy and acceptance criteria. We cut a slice against your spec (your categories, your labels, your format) so you're evaluating conformance, not our defaults.
Send a spec →You write the spec, we fund it to the network. QC-passed, annotated, format-native footage in your hands inside two weeks. Small enough for a corporate card, real enough to train on.
Scope a pilot →Hundreds to thousands of hours to your exact spec, on 30–60 day cycles. Quality risk sits with us: you pay for accepted clips only. Retainers and exclusivity windows available.
Scope an order →The bottleneck in physical-AI data isn't pixels; it's rights. We built the consent chain first and the collection network on top of it.
Every contributor attests IP assignment, likeness and biometric release, and bystander consent before their first frame, with the consent text version recorded against each clip, not assumed across an account.
Every frame face-scanned; detections blurred. Minors auto-rejected. Scan manifests ship with the delivery, so your PII posture is a document you can hand your counsel, not a promise you inherited.
Each accepted clip is hashed with its consent record at acceptance and the hash rides the contributor's on-chain payout memo. Audit any file in your delivery independently: the payment and the provenance are the same record.
There is no shortage of vendors who will point a camera at a kitchen. The question is what arrives in your training loop, and whether you can prove where it came from.
Tell us the task, the volume, and the acceptance criteria, and we respond within one business day with a spec draft and a timeline. Or just ask for the demo slice: it's free, and it ships with the loss curve.