Folding Is Where Garment Data Stops
Ask what garment data exists in the open and the answer is towels and t-shirts being folded on tables. Folding became the benchmark because it is the most tractable garment task: a flat workpiece, a short horizon, and success you can see from one camera. Useful, and nowhere near the skill ceiling of real garment work.
Humanoid labs and manipulation research groups are now chasing that ceiling, because garment handling appears in almost every environment a general-purpose robot is promised for: homes, hotels, hospitals, laundries, retail backrooms, and apparel production. The demonstration data for those tasks has to come from somewhere, and it is not on Hugging Face.
The Operations We Collect
Our capture studios run a task taxonomy built with people who work with garments for a living. Each operation is demonstrated by trained garment operators, not general crowd workers, because the demonstration ceiling is the operator's skill ceiling.
| Operation | What Gets Demonstrated | Why It Is Hard |
|---|---|---|
| Sewing machine operation | Feeding, guiding, and tensioning fabric through a machine; seam start and stop; corner turns | Continuous bimanual control against a moving tool with tight tolerances |
| Ironing | Tool use over a deformable surface, smoothing strategy, garment repositioning between passes | Force control, surface reasoning, and long horizons on changing state |
| Fastening | Buttons, zippers, hooks, and snaps, opened and closed on garments in realistic states | Millimeter-scale bimanual dexterity on soft, occluded targets |
| Hanging | Locating collar and shoulders, inserting hangers, racking and unracking | Cloth state estimation plus insertion under occlusion |
| Dressing forms | Putting garments onto and off mannequins and dress forms | The closest safe proxy for assistive dressing, a major humanoid use case |
| Folding and stacking | Category-specific folds from realistic crumpled starts, stacking and shelving | The baseline task, collected with the initial-state diversity open data lacks |
Coverage Axes That Make It Trainable
- Garment categories. Shirts, trousers, knitwear, outerwear, childrenswear, and linens, each with its own drape and failure modes.
- Materials. Wovens, knits, and technical fabrics across the weight range, because denim and silk are different manipulation problems wearing the same product category.
- Initial states. Fresh from a dryer pile, inside-out, buttoned, twisted. Specified per episode, staged deliberately, never left to chance.
- Recovery behavior. Slipped grasps, misfeeds at the sewing machine, hangers that miss. Operators demonstrate the recovery, and episodes are labeled accordingly.
Capture methods follow the task: teleoperated rigs where on-robot action streams are required, UMI-style handheld grippers and instrumented egocentric capture where human dexterity at full speed is the signal. Everything is collected in purpose-built capture studios in Asia Pacific and delivered as aligned episodes in LeRobot, RLDS, or GR00T-compatible format within 60 days of a signed spec.
Why Operator Skill Is the Product
Anyone can demonstrate folding a towel. Almost nobody outside the apparel trade can feed a curved seam through a sewing machine at consistent tension, or press a shirt in an order that does not re-wrinkle the half already done. Skilled garment work encodes years of tacit knowledge in the trajectories: where to pinch, how much slack to hold, when to reposition. That knowledge is exactly what a policy needs to imitate, and it cannot be faked by giving a crowd worker an iron and a checklist. Our capture studios employ operators with genuine garment production experience, and the difference is visible in the data: smoother tool paths, consistent strategies across episodes, and recoveries that reflect how a professional actually fixes a misfeed.
The deliverable follows the same discipline as every collection we run: synchronized multi-view episodes with aligned action streams, stage and success labels against the agreed taxonomy, calibration files, per-episode QA results, and a data card, shipped in LeRobot, RLDS, or GR00T-compatible format within 60 days of the signed spec.
Who This Data Is For
Humanoid platform teams training household and hospitality skills. Research groups working on deformable manipulation, bimanual coordination, and assistive dressing. Teams building apparel and laundry automation who need pretraining and evaluation data grounded in real operations. If that is you, the wider deformables context is on our deformable manipulation page, and the industrial background lives in the textile and garment robotics guide.
Name the operations and garment categories on your roadmap. We reply with a task taxonomy, coverage plan, and 60-day delivery date. Get a collection quote.