01
Brief
Name the skill, the environment, and what counts as a pass. Set the quality bar and the deadline.
AuraOne · Human Data OS · Robotics
Robots learn from real work. We make it the kind a model can train on.
The open robot datasets total about five thousand hours. A model cannot learn a kitchen from that. We ship real human demonstrations at training scale — reviewed, rights-cleared, and delivered in RLDS, OpenX, or HDF5.
Start with a pilot: 100 to 500 reviewed episodes of one skill, in one environment, on a deadline. A first dataset, not a platform tour.
02
Consent
Attach the rights before the first frame: who recorded it, what license it carries, and a worker-visible receipt they can read.
03
Capture
Operators and teleop reviewers submit demonstrations from signed sites, each clip tied to its device, its review, and the person behind it.
04
Deliver
Accepted episodes ship in RLDS, OpenX, HDF5, or BVH, with a data card, a manifest, and checksums your training stack reads today.
Large-media pipeline
Built for hours of footage, not demo clips.
A robotics program creates gigabytes per worker and terabytes per dataset. We move that footage in resumable segments, prove every object with byte counts and checksums, and stream the accepted clips into the format your stack reads — so what arrives is ready to train, not a folder to clean up.
Segmented capture
Long sessions are organized as bounded segments, so one failed or reworked clip does not spoil the whole assignment.
Direct object-store upload
Raw media moves through resumable multipart upload with signed part refresh instead of making the app server carry multi-GB video.
Integrity proof
Completed objects carry byte counts, checksums, provider proof, request IDs, and explicit integrity status before review or delivery.
Streaming exports
Accepted segments package into robotics datasets with streaming writers, artifact manifests, retention policy, and source lineage.
Cloud delivery
Buyer handoff can use signed links or provider-side cloud copy, with every delivered object tracked by checksum and recall target.
Storage control
Raw, proxy, export, and egress bytes roll up by buyer, program, dataset request, storage class, budget, and lifecycle state.
Dataset packs
Nine packs. Start with the one your robot keeps failing.
Homes, kitchens, laundry, warehouses, retail shelves, cable work. And the two that move a policy: FailureOps captures dropped grasps, blocked views, and recoveries; TeleopOps captures the human corrections behind every intervention. The hard examples, labeled.
HomeOps v1
Dishes, bed-making, trash, counters, cabinets, drawers, and everyday object handling.
CleaningOps v1
Surface wipe-down, clutter pickup, spills, tools, before/after states, and safe recovery.
LaundryOps v1
Sorting, folding, hampers, drawers, hangers, garment variants, and failed-fold recovery.
FoodOps v1
Pouring, stirring, opening containers, appliance use, food prep, and safe cleanup.
WarehouseOps v1
Picking, packing, scanning, shelving, totes, carts, and exception handling.
RetailOps v1
Shelf photos, planogram checks, out-of-stock handling, item movement, and display recovery.
CableOps v1
Plugging, unplugging, cable routing, port alignment, and rack-style manipulation.
FailureOps v1
Drops, spills, failed grasps, blocked views, wrong objects, interruptions, and recoveries — labeled so the same failure is caught next time.
TeleopOps v1
Teleop sessions, human-intervention logs, policy-failure tags, and before/after correction capture — the up-the-curve eval data.
Every scoped pack follows the same review path: rights-cleared episodes in RLDS, OpenX, HDF5, or BVH, each one carrying the consent receipt and review that let it pass. Pilots scope reviewed episode counts before kickoff.
Signed, not pooled
From first frame to final manifest.
Every clip carries who recorded it, what they were asked to do, what rights attach, why it passed, and where it can go. Identity-verified consent, on supply you run, never pooled in one place. Built for the month the largest data vendor was absorbed by one of the labs it served, and a competitor lost four terabytes — including who its workers were.
The EU AI Act enforces training-data provenance in August 2026. Most teams still cannot trace or validate where their data came from. Each datapoint here can.
Task spec01
Consent receipt02
License scope03
Device and sensor metadata04
QA score05
Reject / rework trail06
Annotation and failure labels07
Data card08
Delivery manifest09
Recall / deletion linkage10
Eval / regression-bank output11
Signed environments
Real places, captured with their owner's consent.
Homes, kitchens, warehouses, hotels, and retail floors become capture sites — the rights-cleared access to real environments that is becoming the durable asset. The location owner signs the consent. Privacy zones mask what the robot should not see, PII is redacted on a logged event, and every worker reads the receipt for their own footage.
Homes and apartments
Commercial offices
Warehouses and logistics
Hotels and senior living
Retail and food-service kitchens
Cleaning and home-service networks
Teleoperation review lanes
Expert workflow evals
Route your program
Robot-learning and VLA teams
Build a dataset
Name the skill, the environment, and the format. Get back reviewed, rights-cleared episodes your training stack reads.
Open intakeApproved mobile operators
Run the capture network
One app for operators: find tasks, sign consent, capture to spec, fix rework, and get paid.
Open intakeLocation owners and workforce channels
Turn locations into capture sites
Make a home, kitchen, warehouse, hotel, or retail floor a rights-cleared capture site, with owner consent and privacy zones.
Open intakeExpert workers and eval buyers
Evaluate expert work
Score credentialed experts against rubrics and golden outputs, with provenance on every judgment and pay tied to the result.
Open intake