Expert-demonstrated manipulation data, recovered and graded for policy learning.
We capture skilled human work in real production settings: three synchronized cameras (a head-mounted egocentric view and one on each wrist), frame-aligned to the millisecond, with hand pose recovered per frame. Expert-graded, consented, and collected to your specification.
Three synchronized cameras, head-mounted plus one on each wrist, frame-aligned to the millisecond, with hand pose recovered per frame. An expert garment worker ironing in a working production setting.
- Composite
- 3 cameras
- Views
- Head + both wrists
- Signal
- Recovered hand pose
Real work, recovered as trainable signal.
Contact-rich, deformable manipulation is where simulation and world models remain weakest. We capture it from working professionals doing their actual jobs, across synchronized cameras, and recover it into trainable signal.
Expert demonstration
Working professionals perform real tasks in production settings, captured with a head-mounted egocentric camera and a camera on each wrist, all frame-aligned to the millisecond.
Signal recovery & structuring
Hand and finger pose recovered per frame. Streams time-aligned to the millisecond. Sub-task segmentation, action labels, and success flags applied against a defined task taxonomy.
Expert grading & packaging
Every episode carries an expert-verified quality grade; only graded data ships. Delivered in LeRobot, RLDS, or raw plus HDF5, with a full datasheet and provenance record.
The raw source signal every later layer is recovered from: head plus a camera on each wrist, frame-synced, as an operator sews two fabrics.
- Stage
- Raw capture
- Views
- Head + both wrists
- Sync
- Frame-aligned
What ships in every episode.
Every demonstration ships as time-aligned layers, frame-synced across all cameras.
| Layer | What it gives a policy |
|---|---|
| Egocentric RGB + active perception | Attention and viewpoint, beyond raw pixels |
| Hand & finger pose | The action itself, retargetable to a robot |
| Contact events (annotated) | Expert-labeled grasp and release moments on flagged segments |
| Wrist-camera views (×2) | Close, hand-locked detail the head camera can't see |
| Task structure | Sub-task segmentation, action and affordance labels, success flags |
| Quality grade | Expert review tier per episode; train on the threshold you choose |
| Provenance & consent | Auditable rights and licensing, ready for procurement |
The top two layers made visible: recovered hand keypoints, the action itself, rendered over the egocentric RGB stream as an expert cuts to a line.
- Stream
- Egocentric RGB
- Overlay
- Hand keypoints
- Layer
- The action
Consented at the source. Documented for procurement.
For a policy you'll deploy, where the data came from matters as much as what's in it. Every operator consents to the capture and its use, and every collection is built to carry an auditable provenance and licensing trail, designed for EU / GDPR review and enterprise procurement from the first frame.
Informed, on the record
Operators consent to the capture and its downstream use before recording. Consent scope is tracked per collection.
Auditable provenance
Source, rights, and licensing travel with every episode as metadata, traceable end to end.
Built for scrutiny
Sourcing transparency and data-protection documentation, designed for European compliance review.
Industrial verticals, captured in structured blocks.
Deformable and contact-rich work across industries, collected to task specification. Garment work is available to evaluate today; every other vertical is captured on request, to your task specification and at program scale.
| Vertical | Capability | Status |
|---|---|---|
| Garment & leather-goods work | deformable manipulation (frontier) | AVAILABLE |
| Restaurant / kitchen prep | long-horizon multi-step + deformable food | ON REQUEST → |
| Light assembly & repair | fine bimanual + tool use + kitting | ON REQUEST → |
| Fresh-produce grading & packing | deformable + high object diversity | ON REQUEST → |
| Packing & moving | pick-place, box folding, load-bearing | ON REQUEST → |
| Gold / jewellery manufacturing | fine tool-mediated dexterity | ON REQUEST → |
| FMCG food manufacturing | process, repetitive, controlled | ON REQUEST → |
The textile sample: expert garment work, ready to evaluate.
Multi-camera egocentric demonstrations from working garment-industry professionals. Documented to the Datasheets for Datasets standard, packaged so an ML engineer gets value in under ten minutes.
- Domain
- Industrial garment / leather-goods work
- Task family
- Deformable, bimanual, contact-rich manipulation
- Sample data
- ~16 hours of graded sample episodes
- Program scale
- Collected to spec, on request
- Subjects
- Expert operators, real production settings
- License
- Commercial, non-exclusive by default
A LeRobot-format open teaser is on its way to Hugging Face for free download. The full pilot dataset is available on request today.
Thanks, we've got it.
Marcus follows up from marcus@sovrano.ai within one business day with full pilot access.
Have a specific task, view, or embodiment in mind? Reply to that email and we scope it.
From sample to production program.
Sample
Pull the teaser or request the full pilot. Load it, inspect the signal, see the grading.
Pilot to spec
Tell us the tasks, embodiment, and modalities you need. We collect a scoped block to your specification.
Production program
Scale to a recurring, multi-hundred-hour collection program across any vertical, exclusive if you need it.
Tell us what your policy needs to learn.
We'll send the textile pilot today and scope the capture you actually need.
Request the pilot dataset