Short answer
Laundry folding robot training data covers the whole textile task: pulling one item from a tangled pile, identifying it, finding corners or seams, spreading, folding to a house standard and stacking. Commercial laundries, hotels and apparel fulfillment sites often hold training video and fold standards, and some can host new recording. How cloth behaves in a robot's grippers still has to come from bimanual robot or teleoperation data.
The deployment problem
Folding is a common showcase task in robot learning, and businesses fold textiles at industrial volume. The gap to close is between a demonstration and a production shift. Production folding starts from a cart of warm, sometimes damp, tangled items of several types, each folded to a client's standard at a pace set by the shift.
A deployment team has to decide which parts of that task business data can feed and which need robot data. The answer differs by step, so this page breaks the task down before discussing sources. For the general two-handed requirements behind every step, see bimanual task demonstrations.
Task anatomy
| Step | What happens | Objects and conditions | Typical failure modes |
|---|---|---|---|
| Separate | Pull one item from a pile or cart | Mixed types, warm or damp, tangled | Two items lifted; item snagged |
| Identify and triage | Recognize type, size and side; spot stains or tears | Towels, sheets, pillowcases, uniforms, retail garments | Wrong type; defect missed and folded |
| Find grasp points | Locate corners, hems, shoulders or waistband | Corners buried in crumpled fabric | Middle grasped; same corner twice |
| Spread | Shake out, drag over a table edge or pin flat | Large items exceed one arm span | Twist, wrinkles, item off table |
| Fold | Apply the fold sequence of the house standard | Fold lines, fold count, finished size | Edges misaligned; wrong standard |
| Stack, hang or bag | Count, stack, place on hanger or in polybag | Carts, shelving, hangers, bags | Stack topples; wrong count |
Two variants matter in practice. In many commercial laundries, large flatwork such as sheets is fed into ironing and folding machines, so the human job is corner finding and feeding rather than hand folding. In apparel fulfillment and returns, the job is refolding a garment to a brand's presentation standard, often around a folding board, then bagging and labeling it.
What data helps each part
| Part of the task | Useful data | Usual source | What it cannot show |
|---|---|---|---|
| Separation and grasp-point finding | Overhead and egocentric images of piles with corner and hem keypoints | Commissioned capture at folding stations | How fabric drags and snags when lifted |
| Identification and triage | Images by item type and condition, with reject reasons | Quality photos; commissioned images | Handling of any kind |
| Fold sequence | Step-segmented video of each standard; fold diagrams and target sizes | Training video; fold-standard cards | Tension, slip and drape |
| Spreading and folding control | Dual-arm robot or teleoperation episodes on the same item types | Specialist robot-data teams | Transfer to a different robot; this is the control data |
| Judging success | Photos of accepted and rejected folds against the standard | Supervisor checks; QC records | Why a fold failed |
Published results put the control layer on robot data. The π0 model was trained on data from single-arm, dual-arm and mobile manipulators and demonstrated dexterous tasks including laundry folding (Black et al., 2024). Earlier work showed fine-grained bimanual skills learned from demonstrations collected on low-cost two-arm hardware (Zhao et al., 2023). Simulation is a complement: BEHAVIOR-1K's OmniGibson simulator models deformable objects alongside rigid objects and liquids (Li et al., 2024), but whether a simulated towel matches your linen is something to test, not assume.
Human video earns its place in the perception and task layers: what each item looks like crumpled, which grasp points experienced folders choose, and the order of folds for each standard. For the broader category of soft, shape-changing objects, see deformable object handling.
Capture and labels
For commissioned recording at a folding station, three viewpoints cover most needs: an overhead camera that sees the fabric's full outline, a head-mounted camera that sees the hands and grasp points, and a side view that captures the height and arc of a shake-out. Shake-outs are fast, so set frame rate and shutter speed to keep fabric edges sharp; check blur on a pilot clip before scaling.
# Illustrative only — synthetic label record for one folding episode
episode_id: fold_site008_0217
item: {type: bath_towel, size: 70x140cm, fabric: cotton_terry, state: dry_warm}
fold_standard: FS-HOTEL-B-03 # thirds lengthwise, then half; finished 23x70cm
operator: operator_A12 (experience: 4 yr)
steps:
- {label: separate, start_s: 0.0, end_s: 2.1}
- {label: find_corner, start_s: 2.1, end_s: 4.0, grasp: {left: corner_A, right: corner_B}}
- {label: spread, start_s: 4.0, end_s: 5.6, method: shake_out}
- {label: fold_1, start_s: 5.6, end_s: 8.3}
- {label: fold_2, start_s: 8.3, end_s: 10.2}
- {label: stack, start_s: 10.2, end_s: 11.4, stack_count_after: 7}
outcome: {meets_standard: true, refolds: 0, reject_reason: null}
reference: {diagram: FS-HOTEL-B-03.pdf, photo: FS-HOTEL-B-03_final.jpg}
excluded: robot actions, grip forces, fabric tension (not present in human video)
Record the fold standard as data, with its diagram, finished dimensions and a reference photo. Without it, a model cannot tell a correct fold for one client from a wrong fold for another. Label refolds and rejects rather than trimming them; a supervisor's correction is useful signal for success detection.
Operator skill varies. In robot learning, results depend on demonstration quality; one study compared multi-operator datasets graded by operator quality with a single proficient operator's set (Mandlekar et al., 2021). Record operator experience so you can filter or weight by it.
Supplier types
- Commercial and healthcare linen services. High-volume towels, sheets and garments; flatwork machines plus hand-fold stations; written fold standards per client.
- Hotel on-premises laundries. Fold standards tied to room presentation; the same linen appears in hotel housekeeping room turnover.
- Uniform rental operators. Garment sorting, hanging and bundling.
- Apparel fulfillment and returns processors. Refolding to brand standards, condition grading and bagging; quality photos are common.
Existing assets include training videos on fold standards, fold-standard cards with diagrams, QC reject records and production counts. Training libraries are covered generally on the catalog page for training materials and LMS content. New footage is the usual route for keypoint-ready capture; the site describes a separate physical-world path for new recordings of hands-on work, where each opportunity has its own requirements.
Rights questions are task-specific. Hotel linen often carries client logos, apparel fulfillment may handle unreleased products under client confidentiality, and soiled linen can contain guests' or patients' belongings. Ask who owns the items on camera as well as who owns the footage.
Limits for this task
- No cloth dynamics. Video shows fabric shape, not the tension, friction or cling of damp cotton. Spreading and folding control needs robot data on comparable items, and textile property data such as deformable product properties helps simulation and grasp planning.
- Human technique. Folders pin fabric with forearms, use the table edge and pinch with whole hands. Some of this does not map to a humanoid's hands; see the human-to-humanoid embodiment gap.
- Speed. Experienced folders work faster than any robot will at first, and shake-outs blur. Working pace in demonstrations covers the trade-off.
- Item variety. Weight, size, pile height and weave differ across clients; item attributes are covered in apparel and textile attributes.
- Happy-path footage. Training video shows clean folds. Tangles, inside-out garments and stained items appear only if collection asks for them.
Questions to ask suppliers
- Which item types are folded by hand, and which go through feeders or folding machines?
- Are fold standards documented with diagrams, finished dimensions and reference photos, and do they differ by client?
- Can recording cover the pile or cart stage, not only folding on a clear table?
- What share of items arrives damp, tangled, inside out or stained?
- Are refolds and rejects logged, and can they be linked to recorded episodes?
- How many operators, at what experience levels, would appear in the footage?
- Do client logos, unreleased products or customer belongings appear on camera?
- Do workers' releases cover recording for AI training and licensing to a third party?
Annotators can segment these episodes to the step list above using the rules in step segmentation annotation.
Discuss your requirements
If you are scoping a folding deployment, describe the item types, fold standards and capture you need to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. It reviews each requirement and follows up about potential sources; a request does not guarantee a match. More task pages sit on the workplace task demonstration hub, and the humanoid training data guide maps the other data types.
Describe the workplace demonstration data your robotics project needs
Sources
- π0: A Vision-Language-Action Flow Model for General Robot Control — laundry folding by a policy trained on robot data, including dual-arm data.
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware — bimanual skills learned from robot demonstrations.
- BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation — simulation of deformable objects for everyday activities.
- What Matters in Learning from Offline Human Demonstrations for Robot Manipulation — demonstration quality and operator variation.