Garments, towels, sheets, bags. Folding, flattening, sorting, hanging — collected as human egocentric demonstrations and bimanual robot teleoperation on the same tasks, with the material properties of every item measured and recorded.
Cloth changes shape with every touch. It self-occludes, its behavior depends on the fabric, and no two crumpled starting states are alike. Robots that handle rigid objects well still struggle here — and the training data to fix it is scarce in ways generic collection doesn't address.
Open egocentric datasets include folding clips — but a policy can't tell silk from denim by pixels alone, and the datasets don't say either. Fabric properties drive cloth behavior, and they go unrecorded.
Human demonstrations teach the skill; robot episodes ground it in an embodiment. Most sources offer one or the other. Adapting a model needs both, on the same tasks, captured with matching camera viewpoints.
A dropped corner, a wrong grasp, a re-grasp and recovery — that's most of what real cloth handling looks like, and it's what policies need to see. Curated-success datasets edit it out.
A garment registry accompanies every dataset: composition from the care label, measured weight class, stiffness and stretch classes from a documented bench test, and the starting state of every episode. Filter, ablate, and diagnose by fabric — not just by task.
The same task battery, captured two ways: human egocentric demonstrations (head-mounted RGB-D with pose, dual wrist cameras) and bimanual robot teleoperation with logged joint states — delivered in LeRobot format, ready to load.
Each delivery ships with its own measurements: cross-camera sync error, hand-tracking coverage per episode, and human-verified segment annotations checked against a gold-standard subset. You see the numbers before you train on the data.
Configured to match how current vision-language-action models are trained: egocentric head view, close-up wrist views, depth, pose, and language labels.
| Stream | What it captures | Format |
|---|---|---|
| Head camera | Egocentric RGB + metric LiDAR depth + 6-DoF camera pose | MP4 · depth maps · pose JSON |
| Wrist cameras ×2 | Close-range view of each hand's contact with the fabric | MP4, synced |
| Hand tracking | Per-frame hand keypoints with confidence and coverage stats | NPZ (T×21×3) |
| Glove hand capture | Sensor-measured finger articulation, immune to visual occlusion — tiers from per-finger flexion to full ~25-DoF hand models, matched to your fidelity requirement. Available for pilot engagements on request | joint angles · NPZ, synced |
| Robot episodes | Bimanual teleoperation: joint states, gripper state, synced cameras | LeRobot V3 · HDF5 · Zarr |
| Annotations | Time-segmented subtask labels, human-verified | JSON |
| Registry & report | Garment properties, episode manifest, per-delivery quality measurements | JSON · Parquet · MD |
⟶ Marked items are the cases shipping folding robots publicly struggle with — we collect them on purpose, including deliberate failure and recovery episodes.
We're a small team and we'd rather show than tell. Request the sample pack and evaluate it against your own pipeline.
REQUEST · SAMPLE PACK
Email a line about what you're building — company or lab — and the tasks you care about. We reply within a day.
info@dextridata.comPilot batches to your spec follow, if the sample holds up.
The same rig, operators, and quality process run any manipulation collection. If your tasks aren't cloth, tell us what they are.
Deformable linear objects — routing, taping, connector work. Harness assembly remains overwhelmingly manual, and demonstration data is nearly nonexistent. Our registry approach maps directly: gauge, stiffness, connector spec per item.
Poly bags, pouches, soft parcels — the warehouse version of the cloth problem. Same rig, same annotation and quality process.
Kitchen and household tasks, pick-and-place, tool use, bimanual coordination — custom human-demonstration and teleoperation collection against your task list and output format.
Yes. We contribute sample batches to open dataset efforts and collect to spec for academic groups working on deformable manipulation — same rig, same registry, same quality report. If you're part of a consortium or lab and a deformable-focused slice would help your coverage, write to us.
No. Deformable objects are the specialization because that's where the data gaps are hardest and our material-registry approach adds the most — but the rig, operator team, and quality process run any manipulation collection. Cable and packaging work uses the same playbook, and general manipulation tasks are collected to custom spec.
Teams training manipulation policies or vision-language-action models that need deformable-object coverage: robotics companies shipping folding or laundry products, humanoid programs adding garment skills, and research labs working on cloth manipulation.
A small set of garment-manipulation episodes across both capture modes — human egocentric RGB-D and bimanual robot teleoperation — plus the garment registry, the episode manifest, and the quality report for that batch. Enough to run through your own loaders and judge for yourself.
Yes — custom collection is the core service. Tell us the tasks, garment types, starting states, and output format; we align on the spec before collection starts. Pilot batches come first so you can evaluate quality before committing to volume.
LeRobot V3, HDF5, and Zarr for robot episodes; MP4 + depth + JSON for egocentric data; Parquet manifests. If your pipeline needs a different layout, we align on it during the spec call.
Every delivery includes its own measurements rather than blanket claims: cross-camera synchronization error, per-episode hand-tracking coverage, and segment annotations verified by humans and checked against a gold-standard subset. The methodology ships with the report.
Based in the Bay Area, CA and Austin, TX, with collection operations in India, run by a trained in-house team — not crowdsourced. The founding team's background is in ML engineering.
No — we're the data layer. We work alongside your hardware and training stack, not instead of it.